Concept Drift Adaptation in Text Stream Mining Settings: A Comprehensive Review

Feature drift considers the changes in the importance of features, signifying that, over time, a subset of features may become necessary for an ML model, while other subsets may become obsolete (Barddal et al., 2017). It constitutes a challenge for an ML model because incrementally defining the best feature set over time can be complex. In addition, the dependent ML model can have its performance degraded over time if the selected feature set is inadequate.

Considering the subcategories of feature drift depicted in Fig. 5, we describe the Changes in important words, Changes in important dimensions, and Vocabulary shift. Different types of features may be considered in ML approaches regarding text-related tasks. For instance, one approach is to consider the texts split into tokens or use direct techniques such as bag-of-words or TF-IDF. These techniques resort to counting tokens and measuring their overall importance, respectively.

Other approaches regard the numerical transformation of texts, such as Word2Vec (Mikolov et al., 2013a). Because the former can be directly related to specific words (or tokens), the changes in those words are regarded in this study as Changes in important words. However, changes in numerical representations without direct relation to words or tokens are regarded as Changes in important dimensions.

Finally, we consider vocabulary shift as one type of feature drift. Vocabulary shift (Amba Hombaiah et al., 2021) ponders the changes of words in a vocabulary maintained by the approach as a type of text drift. Different from the aforementioned subtypes of feature drift, vocabulary shift considers the changes, i.e., addition or removal of items, in the internal structure that stores the tokens. Amba Hombaiah et al. (2021) compared vocabularies in year-timed slices, measuring changes between vocabularies from different years.

Melidis et al. (2018), Chamby-Diaz et al. (2019), and Amba Hombaiah et al. (2021) addressed one of these aforementioned categories of feature drift directly. Melidis et al. (2018) proposed an ensemble-based method for predicting feature values in the next time point. Considering this case, the work is categorized as Changes in important words because their method uses a sketching mechanism to retain essential words in a fixed-size feature space, according to their occurrence count. In one version the authors presented, they utilized ADWIN (Bifet and Gavalda, 2007) to evaluate a significant decrease in word usage to decide when to remove it from the sketch.

Chamby-Diaz et al. (2019) proposed a feature selection method based on correlation suitable for data streams, categorized as Changes in important dimensions. Although the method was not developed specifically for use on text streams, the authors demonstrated its use on a text-related dataset, i.e., a spam dataset. Their method retains a covariance matrix coupled to a concept drift detector. Whenever it receives a warning signal, the covariance matrix is incrementally updated. When the concept drift detector triggers a drift signal, a one-pass algorithm computes feature-feature and feature-class correlations. Subsequently, a new Naive Bayes model is trained based on the new feature subset, which is chosen according to the merit of each feature subset from the correlation-based feature selection method (CFS) (Hall, 1999).

Unlike prior works, Amba Hombaiah et al. (2021) used vocabulary shift to estimate the changes in the usage of tokens across several years, i.e., between 2013 and 2019. The authors proposed sampling methods for updating BERT models (Devlin et al., 2018) to maintain the models’ usefulness in text-streaming scenarios. Initially, the authors emphasized that “vocabulary is the foundation of language models”. However, vocabularies can contain different types of representation, such as complete words and sub-word segments, e.g., wordpiece (Devlin et al., 2018). The authors analyzed the vocabulary shift considering the 40,000 most frequent tokens, accounting for hashtags, and wordpieces. Regarding hashtags in the years 2013 and 2019, the vocabulary shift was 78.31%, while for wordpieces in the same period, the shift was 38.47%. The authors argued that these results and their analysis justify the development of such an incremental method proposed by them. Furthermore, the authors stated that although larger vocabularies may lessen the vocabulary shift, they are more computationally costly and, therefore, potentially infeasible for real-world scenarios.

We consider real drift according to the definition in Gama et al. (2014), which is changes in $p(y|X)$ that can occur with or without changes in $p(X)$. Considering this case, $X$ regards the input features, while $y$ corresponds to the class, and $p$ is the probability. Real drift in a classification task refers to the change in the classes’ boundaries, which may be accompanied by changes in the data distribution in $X$. In this work, few papers handle different types of real concept drift, e.g., sentiment drift. However, because they regarded changes in $p(y|X)$, these papers were categorized as real drift.

This study considers topic drifts as an extension of real drifts. In the literature, topic drifts are encountered in applications regarding topic modeling, topic labeling, and short-text classification. Thus, a topic could drift by the change of either text labeled as a particular topic, i.e., $p(y|X)$), or by the change of a topic distribution in the stream, i.e., $p(X)$, or both simultaneously. In addition, it is common to use methods based on Latent Dirichlet Allocation (LDA) in short-text-related applications.

A significant number of papers (Pohl et al., 2018; Abid et al., 2018; He et al., 2018; Suprem and Pu, 2019a; Suprem et al., 2019b; Wang et al., 2019; D’Andrea et al., 2019; Mohawesh et al., 2021; Heusinger et al., 2020a; de Moraes and Gradvohl, 2021; Bechini et al., 2021; Sun et al., 2021; Amba Hombaiah et al., 2021; de Mello et al., 2018; Heusinger et al., 2022; Rabiu et al., 2022; Bravo-Marquez et al., 2022; Bondielli et al., 2022; Kolajo et al., 2022; Fenza et al., 2023; Susi and Shanthi, 2023; Assenmacher and Trautmann, 2022) regard exclusively real drifts. Most commonly, methods in this category either: (i) use concept drift detectors to detect drift and trigger the model update or (ii) update the model regularly.

Suprem and Pu (2019a), Suprem et al. (2019b), and Suprem and Pu (2019b) presented from multiple perspectives a system for detecting physical events with emphasis on landslides, i.e., the sudden mass of rock and earth movements downwards steep slopes. They combined data from social media (which is voluminous but not so trustworthy) and governmental reports (scarce but trustworthy) to train a model for landslide detection. The authors argued that the terminology landslide can suffer concept drift because of its use in different contexts, such as politics. In their case, the model is updated regularly, using the governmental reports as ground truth. However, Mohawesh et al. (2021) and Heusinger et al. (2022) utilize concept drift detectors to detect drifts explicitly. Mohawesh et al. (2021) used ADWIN (Bifet and Gavalda, 2007), DDM (Gama et al., 2004), EDDM (Baena-Garcıa et al., 2006), and Page Hinkley (Page, 1954; Sebastião and Fernandes, 2017), while evaluating fake reviews detection. The authors claimed that fake reviews could lead customers to make poor decisions. Also, it is an adversarial problem: once models become better at detecting fake reviews, the unlawful reviewers change patterns over time to overcome the models. The adversarial aspect of this problem results in concept drift, which can cause the models’ performance to degrade over time.

Susi and Shanthi (2023) proposed a complete system for tweet collection, automated training data generation, and BERT (re)training for sentiment prediction and adaptation to sentiment drift, namely Twitter Sentiment Drift Analysis - BERT (TSDA-BERT). The authors used Apache Kafka to simulate the Twitter stream. A BERT model has on top a three-layer dense network that performs the classification. Since the sentiment drift is verified using the predictions, we categorized this paper in the real drift category.

Assenmacher and Trautmann (2022) proposed a 2-phase online method for textual clustering, namely textClust. This method leverages TF-IDF to decide the proximity of incoming text to microclusters. In addition, the authors take advantage of unigram and bigram representations and use cosine similarity to evaluate the most suitable cluster to include the incoming text when possible. Over time, in the offline phase, the method can maintain the model concise by merging similar clusters and removing outdated ones. To define the outdated clusters, the authors used a fading factor for the cluster weights. The authors mentioned that the fading factor helps the model handle concept drift.

Another significant number of papers (Murena et al., 2018; Li et al., 2018; Yin et al., 2018; Hu et al., 2018; Hammer and Yazidi, 2018; Rakib et al., 2021; Yang et al., 2021; Van Linh et al., 2022; Nguyen et al., 2022; Vo, 2022; Li et al., 2022) approach the Topic drift problem. Topic drift primarily refers to short-text-related tasks, which commonly require additional steps to provide satisfying results, e.g., data enrichment step or use of statistical information of the application context. Li et al. (2018) proposed a method for short-text classification using feature space extension. Probase (Wu et al., 2012), an open semantic network, is used for the extension. According to Li et al. (2018), Probase was selected by the availability of several super-concepts. It means that, in order to enrich a short text, they could obtain more information from Probase, e.g., super-concepts(Apple) = [company, tech giant, large company, manufacturer], and add it to the short text. Rakib et al. (2021) developed EStream, a method for efficient short-text clustering. Their approach used lexical, e.g., bigrams, unigrams, biterms, and semantic information from GloVe (Pennington et al., 2014) to define the clusters. Changes in proximity between text and clusters over time are used to determine whether a concept drift occurred.

Both Murena et al. (2018) and Hu et al. (2018) used LDA (Blei et al., 2003) to address their challenges (short-text classification and topic modeling, respectively). As Li et al. (2018), Hu et al. (2018) enriched data using external sources. They employed LDA to mine hidden information from these external sources to add the top representative words in the short texts. Drifts are flagged by calculating the semantic distance between each short text in the current and subsequent chunks. Similarly to Hu et al. (2018), Murena et al. (2018) used LDA for topic modeling in document streams. The authors, in this case, integrated an ADWIN to LDA to detect topic drifts.

Li et al. (2022) presented a method for short-text classification in text stream scenarios. The authors enriched short texts by using representations from BERT and Word2Vec. Both were trained using massive corpora, which, according to the authors, should be highly consistent with the topics related to the datasets the authors evaluated. In addition, the authors proposed a distributed LSTM-based ensemble method that includes a concept drift factor. The concept drift factor is used to determine the importance of an LSTM layer in the final result.

Semantic shift regards changes in the meaning of tokens over time. It is most commonly handled in papers that study linguistic changes over several years, decades, or even centuries. Generally, the datasets that support these tasks are entitled diachronic. However, semantic changes can also occur within a short time, such as in weeks (Stewart et al., 2017). The semantic shift was briefly introduced and discussed in Section 2.

Amba Hombaiah et al. (2021), Lu et al. (2022), and Periti et al. (2022) approach the problem of semantic shift. Amba Hombaiah et al. (2021) discussed the semantic shift as an analysis of whether it occurred. In the specific task of hashtag prediction, the authors evaluated the shift in top contextual words of the hashtags #china, #uk, and #usa, considering the years 2014 and 2017. The authors agreed that, in 2014, the contextual words related to #usa were related to the World Cup, while in 2017, the words were related to US politics. However, Periti et al. (2022) aimed at detecting semantic shifts incrementally. In this case, the authors applied clustering methods, such as affinity propagation, to generate clusters in time slices. The authors determined a semantic shift by measuring the distance between embedding sets using metrics such as Jensen-Shannon divergence (Nielsen, 2019) and the distance between prototype embeddings.

Lu et al. (2022) presented a word-level graph-based method to generate dynamic word embeddings. The fundamental concepts are around maintaining long-term and short-term word-level knowledge graphs. These graphs preserve the co-occurrence between words. The relations between words help define the occurrence of semantic shifts. For semantic shift detection, the authors evaluated the closest words to apple (in the New York Times dataset) and network (in the Arxiv dataset). In addition, the authors evaluated their method by considering trend detection and text stream classification. In both (Lu et al., 2022) and (Periti et al., 2022), the words of interest in detecting semantic shifts must be known in advance.

According to Gama et al. (2014), virtual drift regards changes in data distribution without changing the boundaries between classes. Using a similar notation as in Section 4.2.2, virtual drift happens when $p(X)$ changes but $p(y|X)$ does not. In addition, Gama et al. (2014) state that different definitions exist for virtual drift in the literature. Papers Suprem and Pu (2019b); Rabiu et al. (2022) and Rabiu et al. (2023) illustrate this category. Virtual drifts must be tracked, particularly in cases where no classes or clusters’ labels $y$ are available.

In Suprem and Pu (2019b), the authors proposed a method for landslide detection. The method relies on social media data and governmental reports. Section 4.2.2 already cited this paper together with Suprem and Pu (2019a) and Suprem et al. (2019b). However, Suprem and Pu (2019b) explicitly emphasized their concern about handling the virtual drift problem. They highlighted that model fine-tuning is sufficient in this case, compared to model re-creation. Nonetheless, no reason for their concern about virtual drifts is provided. Rabiu et al. (2022) presented a two-component method for concept drift detection applied to sentiment analysis and opinion mining. Similar to Suprem and Pu (2019b), Rabiu et al. (2022, 2023) handle virtual drift. Although it is not explicit in the papers, the drift detection method uses two windows to evaluate possible concept drift based on a distance metric to be selected. Different from most works that couple a concept drift detector with a classifier to utilize the classification errors as a proxy for the detector, Rabiu et al. (2022, 2023) used the input data, thereby using the concept drift detector to check $p(X)$.

Adaptive corresponds to a self-updating model without explicitly detecting drift but rather from time to time. This category is called blind adaptation in Gama et al. (2014). A substantial number of papers (Pohl et al., 2018; Abid et al., 2018; Melidis et al., 2018; Yin et al., 2018; He et al., 2018; Abid et al., 2019; Suprem and Pu, 2019a; Suprem et al., 2019b; D’Andrea et al., 2019; Heusinger et al., 2020a; de Moraes and Gradvohl, 2021; Bechini et al., 2021; Amba Hombaiah et al., 2021; Rakib et al., 2021; Yang et al., 2021; Van Linh et al., 2022; Nguyen et al., 2022; Vo, 2022; Lu et al., 2022; Bondielli et al., 2022; Kolajo et al., 2022; Assenmacher and Trautmann, 2022) consider adaptive approaches. Pohl et al. (2018) proposed a batch-based method with an application for crisis management in social media. The method is based on active learning and queries a user whenever the classifier fails to confidently determine whether the input text is relevant to the task. The authors selected two events corresponding to two subsets from a more extensive dataset, i.e., Colorado floods and Australian bushfires, and 1000 data points were labeled via crowd-sourcing. In addition, the authors mentioned that labeling data is particularly costly in streaming scenarios; however, it still requires a human in the loop in a task such as crisis management. According to the authors, this model can adapt itself in case of concept drift using the characteristics of the ML technique. For example, although they applied their scheme using k-Nearest Neighbors (k-NN) and Support Vector Machines (SVM), it could be any other classifier. For instance, the authors claimed that when using k-NN and SVM, the continuous calculation of the boundaries results in drift adaptation.

Amba Hombaiah et al. (2021) split the social media data by considering the years of publication. The work considers two datasets, corresponding to three different tasks: (i) 2014 Country hashtag prediction, (ii) 2017 Country hashtag prediction, and (iii) OffensEval 2019. The authors compared seven methods for each scenario: two static BERT models and five dynamic BERT models. Considering the static BERT models, one is trained with data from the previous year, and the other uses data from the current year. For example, considering the 2014 Country Hashtag prediction task, one model is trained with tweets from 2013, and the other (a model checkpoint from the first model) is updated with an amount of data from 2014. The dynamic BERT models are fine-tuned using sampled tweets from the current year using different sampling methods, e.g., uniform random, weighted random, token embedding, sentence embedding, and token masked language modeling (MLM) loss. The sampling methods define different strategies for the model to overcome drifts/semantic shifts over time. The uniform random sampling method regards sampling methods in which the tweets from the current year are sampled randomly. In addition, the weighted random regards a sampling method in which the tweets from the current year are sampled randomly, considering the number of wordpieces generated by the tokens in the current year’s tweets. However, token embedding, sentence embedding, and token MLM loss differ. The token embedding method assigns higher weights to tweets that contain new tokens and random samples from the current year’s tweets. The sentence embedding method calculates the cosine distance between the updated and the current models. Both cosine distance and tweet length are used to determine a score and then perform the sampling. The token MLM loss method considers the last layer from the BERT model, masks out 15% of the tokens and uses the surrounding words to predict the masked ones. A high loss value may indicate drifts.

Yin et al. (2018) proposed two algorithms for short-text stream clustering: MStream and MStreamF, a concise version that deletes outdated clusters. The algorithms receive document batches and are one-pass, in which the first document creates a new cluster, and the subsequent either selects one of the clusters to be assigned to or creates a new cluster. This assignment occurs after the batch is processed. The authors argued that concept drift is handled by assuming that the documents are generated by a Dirichlet Process Multinomial Mixture (DPMM) (Antoniak, 1974) and thus derive the probabilities of documents belonging to existing clusters.

D’Andrea et al. (2019), Bechini et al. (2021), and Bondielli et al. (2022) tackled two different problems similarly: stance detection about vaccination and the Green Pass (as the EU Digital COVID Certificate is known), both in Italy. The authors in both works categorized the application into stance detection, a branch of sentiment analysis. In these cases, the tweets are classified in a three-class fashion as either (i) in favor, (ii) neutral, or (iii) not in favor. D’Andrea et al. (2019) and Bechini et al. (2021) analyzed public opinion about vaccines in Italy based on tweets. D’Andrea et al. (2019) addressed concept drift by incrementally retraining the model, such as an SVM model. However, they emphasized that, considering their dataset, incremental retraining could not outperform a static SVM regarding accuracy. In Bechini et al. (2021), concept drifts are handled similarly to D’Andrea et al. (2019). However, the tweets from the new batch are semantically weighted according to previous events. Thus, the authors reached better values than other approaches, e.g., static model, regular retrain, DARK (Costa et al., 2017), and the proposed semantic scheme. Although Bechini et al. (2021) was published in 2021, it was applied to regular vaccination, unrelated to Covid-19. However, Bondielli et al. (2022) covered the opinion about the Green Pass concerning Covid-19. The authors evaluated different schemes to handle concept drift, including retraining with sliding windows and an ensemble of classifiers. The complete retraining led to the best average accuracy. Still, the highest feature space was reached due to the data accumulation and the utilization of TF-IDF as a text encoding method that generates a very high-dimensional representation.

Assenmacher and Trautmann (2022) presented an online method for textual clustering, i.e., textClust. In order to overcome concept drifts, the method leverages a fading factor. It helps the model to exclude stale clusters. In addition, there is another parameter $tr$ that dynamically determines the distance limit for a cluster to merge with another. This is also used to help determine whether a new input instance should be incorporated into a given cluster.

Explicit approaches directly detect the drift, via statistical tests or distance calculation. As examples of statistical tests used in the selected papers, we mention the Page-Hinkley test (Page, 1954; Sebastião and Fernandes, 2017), and ADWIN (Bifet and Gavalda, 2007). As examples of distance calculation metrics, we cite the Jensen-Shannon divergence (Nielsen, 2019), the Kullback-Leibler divergence test (Kullback and Leibler, 1951), and the cosine distance. Approaches presented in the papers Murena et al. (2018); de Mello et al. (2018); Li et al. (2018); Hu et al. (2018); Hammer and Yazidi (2018); Chamby-Diaz et al. (2019); Suprem and Pu (2019b); Wang et al. (2019); Mohawesh et al. (2021); Sun et al. (2021); Heusinger et al. (2022); Rabiu et al. (2022); Periti et al. (2022); Li et al. (2022); Fenza et al. (2023); Susi and Shanthi (2023); Rabiu et al. (2023) explicitly handle concept drift.

Concerning explicit detection using statistical tests, Mohawesh et al. (2021) tested four concept drift detectors: ADWIN (Bifet and Gavalda, 2007), DDM (Gama et al., 2004), EDDM (Baena-Garcıa et al., 2006), and Page-Hinkley test (Page, 1954; Sebastião and Fernandes, 2017). We considered ADWIN as a statistical test because, in the original paper, the authors indicated that their statistical test verifies whether the observed average in subwindows is above a defined threshold (Bifet and Gavalda, 2007). In addition, DDM (Gama et al., 2004) and EDDM (Baena-Garcıa et al., 2006) perform evaluations based on the statistical properties of a stream and thus are considered in this work a statistical test. Mohawesh et al. (2021) simulated concept drift by splitting the temporally ordered dataset into five chunks and rearranging them. The concept drift detectors use the calculated accuracy over the most recent input data as a proxy, i.e., a window size of 200. ADWIN and EDDM had the best accuracy (coupled with a classifier) among the scenarios tested in the study.

Heusinger et al. (2022) proposed a method that uses random projection for dimensionality reduction using text streams as input. In their experiments, preprocessing was done offline for the whole dataset to generate TF-IDF and embedding representations. Thus, their process is not fully incremental, except for the dimensionality reduction method, which is incremental (in batches). Considering the real-world dataset, i.e., NSDQ, proposed in the same paper, the authors obtained a vector representation of 3442 dimensions using TF-IDF. Using their online dimensionality reduction method, NSDQ was projected onto 200 dimensions. The authors concluded that random projection could reduce the run time, even considering the offline preprocessing time. To detect concept drift, the authors used KSWIN (Raab et al., 2020), based on the Kolmogorov-Smirnov test (Kolmogorov, 1933; Smirnov, 1948). In this case, KSWIN monitors every dimension of the vector representation. In addition, the authors mentioned that different types of concept drift might be present because NSDQ is a real-world dataset (Heusinger et al., 2022). Their assessment of concept drift detection relies on true positives and false positives. However, it is unclear how both metrics were calculated due to the absence of labeled drifts in the dataset. The results indicated more concept drifts detected in the original space, an expected outcome because KSWIN monitors each dimension separately. Finally, the authors mentioned that models trained with original and projected feature spaces maintain the same level of accuracy. Both Suprem and Pu (2019b) and Heusinger et al. (2022) used t-SNE (van der Maaten and Hinton, 2008) plots to support the existence of concept drift in the datasets on which they applied their proposed methods.

Considering the Explicit detection with the aid of distance metrics, Li et al. (2018) developed a method for short-text classification in the presence of topic drifts. As explained in Section 4.2, the approach automatically enriches the short texts by using Probase. The topic drift detection is performed as follows: the short-text stream is received in chunks, and after they are clustered, the label distribution can be evaluated using the clusters. Subsequently, the distance between the cluster centers in sequential chunks is calculated using the cosine distance. According to the value obtained, the method categorizes it either into: (a) no drift; (b) noisy impact; or (c) topic drift. In addition, the authors simulated topic drifts by generating datasets with topic changes after fixed periods. Their detection method was compared to nine drift detectors. In terms of false alarms, missing drifts, and delay the proposed method obtained high average rankings, being statistically equivalent (using the Bonferroni-Dunn test) to the best drift detectors in each metric.

Rabiu et al. (2022, 2023) developed an ensemble classifier coupled to a novel mechanism for drift detection-based adaptive windows (DDAW). Their method suits text streams, especially users’ sentiments and opinions. Their approach can be divided into two components: (i) drift detection and (ii) classification. In many applications, classification errors are used as a proxy for the drift detector. However, the drift detection component compares the data distribution considering two windows. Thus, it is possible to measure drift by evaluating the dissimilarity between the windows. An intriguing aspect of this approach is that it allows for distance metrics and statistical tests. In the paper, the authors compared the Hellinger distance (Hellinger, 1909), Kullback-Leibler divergence (Kullback and Leibler, 1951), Total Variation distance, and the Kolmogorov-Smirnov test (Kolmogorov, 1933; Smirnov, 1948). Their approach, coupled with the Hellinger distance, obtained the best values regarding false alarms, detection rate, and accuracy, even compared to other drift detection methods, i.e., AEE (Kolter and Maloof, 2005), RDDM (Barros et al., 2017), and Page-Hinkley (Page, 1954; Sebastião and Fernandes, 2017). It is unmentioned how the drifts are labeled or whether the data is rearranged to simulate drifts.

Suprem and Pu (2019b) developed a system for landslide detection, a physical event that causes destruction and for which there are no physical sensors to detect. The authors combined data from social media and governmental agencies to perform the detection. Concept drift is detected using the Kullback-Leibler divergence test (Kullback and Leibler, 1951) to evaluate the distribution of two batches. The model is updated by generating or updating the classifiers to handle the concept drift.

Li et al. (2022) presented a distributed long short-term memory (LSTM)-based ensemble method for short-text classification in text stream scenarios. The short texts are enriched by using BERT and Word2Vec models. The LSTM-based method includes a concept drift factor used as a threshold to compare the distance between the LSTM layer trained with the previous batch and the layer trained with the current batch. If the concept drift factor is above the threshold, the weight of the current layer will be higher to generate the combined final output.

Fenza et al. (2023) proposed a fuzzy-formal-concept-analysis-based index for concept drift detection and applied the method to a fake news classification problem. Although the concept drift detection is not directly approached, the authors calculated the correlation between the classifier’s performance and the proposed index. The index is calculated from a fuzzy lattice, i.e., a fuzzy hierarchical knowledge structure, while the classifier’s performance is calculated using F-Score and accuracy. Their results demonstrated a high (Sperman’s and Person’s) correlation, between 69% and 87%. The authors claimed that the method has the potential to be used as a proxy for the model update process. In addition, the fuzzy lattice seems never to be updated, which may hamper the model from properly working over a long time.

Susi and Shanthi (2023) proposed a sentiment drift analysis system based on BERT models, namely TSDA-BERT. According to the authors, the system receives data in a sliding window fashion corresponding to four days. The authors calculated the positive and negative scores per window based on the proportion of them in the window. From these values, a sentiment drift measure is calculated by simply subtracting the number of negative from the number of positive tweets. This measure is used for sentiment drift detection by calculating it between time periods; if the score is negative and goes positive or vice-versa, it indicates a drift.

In this category, the works propose techniques that create, update, and combine multiple models, the so-called ensembles. Over time, an ensemble can be updated by removing outdated base learners while adding new base learners trained on newly arrived data. In Suprem et al. (2019b), the presented system for landslide detection uses batches to update the model. The landslide detector uses a classifier, which is an ensemble. The authors mentioned that they used two approaches for selecting base learners: relevancy and recency. When using relevancy, a k-NN search is performed to discover the most relevant base learners from a pool of trained base learners, considering the centroid of the data used to train these learners. However, the recency scheme returns the most recent base learners used to compound the ensemble. In addition, the weighting scheme can be configured as an unweighted, weighted, or model-weighted average. The unweighted average equally considers the base learns to provide an output. The weighted average considers weights provided by domain experts, and the model-weighted scheme considers the base learners’ prior performance to weigh them.

Sun et al. (2021) described an ensemble classification model for short text classification in environments bound to concept drift. The paper emphasizes three main aspects: a feature extension based on the short text features, a concept drift detection method, and an ensemble model. Considering the ensemble model, the authors used SVM as a base classifier. A new classifier is added when concept drift is detected. If the classifier pool is complete, the oldest classifier is removed to add the current one after being trained on the new batch.

Hu et al. (2018) proposed a short text stream classification method based on content expansion coupled with a concept drift detector. The expansion was performed by adding information from external sources, and 100 Wikipedia pages related to 50 keywords were selected, totaling 60,600 pages. The classification task in this study was performed using an ensemble of SVMs, in which each base learner is trained per chunk using the expanded texts. The number of base learners is limited to a specific parameter $H$: when this number is met, the oldest learner is replaced. In specific situations, the latest learner can replace an older learner trained using semantically similar chunks.

Rabiu et al. (2022, 2023) presented an ensemble method for classification. Particularly, Rabiu et al. (2023) tackles the sentiment classification problem. The ensemble model is updated over time by removing the worst base learner from the ensemble when it reaches the maximum number of base learners. To determine the worst base learner, a weighting calculation is performed by leveraging the base learner’s mean square error on the new input data, i.e., $\textrm{MSE}_{i}$, and the base learner’s mean square error on the data from the previous batch (reference data), i.e., MSE${}_{r}$. The complete weight calculation for a base learner is performed as $weight=\frac{1}{\textrm{MSE}_{r}+\textrm{MSE}_{i}+\alpha}$, where $\alpha$ is a non-zero factor to avoid division by zero.

The Incremental update scheme regards models capable of learning from new pieces of data without completely retraining the model. In our selection, several papers (Murena et al., 2018; Abid et al., 2018; Li et al., 2018; Melidis et al., 2018; Yin et al., 2018; Hammer and Yazidi, 2018; He et al., 2018; Abid et al., 2019; Suprem and Pu, 2019a, b; Mohawesh et al., 2021; Heusinger et al., 2020a; de Moraes and Gradvohl, 2021; Rakib et al., 2021; Yang et al., 2021; Van Linh et al., 2022; Nguyen et al., 2022; Heusinger et al., 2022; Vo, 2022; Lu et al., 2022; Bravo-Marquez et al., 2022; Kolajo et al., 2022; Periti et al., 2022; Assenmacher and Trautmann, 2022) employ incremental models to approach their applications. However, we distinguish between the manners in which the data are inputted into the model: (i) One input at a time; and (ii) In batches.

Heusinger et al. (2020a) proposed a method for dimensionality reduction using random projection. As already cited in Section 4.3.2, the process is not fully incremental. In this study, the authors utilized three classifiers: (i) Adaptive Robust Soft Learning Vector Quantization (Heusinger et al., 2020b), (ii) Adaptive Random Forest (Gomes et al., 2017), and (iii) Self-adjusting Memory k-NN (Losing et al., 2017). The dimensionality reduction method uses a window of size 1000. However, when applied to the classification methods, the process in incremental One input at time, except for the Self-adjusting Memory k-NN, which Heusinger et al. (2020a) cited that they used as parameters five neighbors and a window size of 1000 to match the window size of the random projection. The authors in Mohawesh et al. (2021) incrementally updated the models. In their case, they used Stochastic Gradient Descent for Support Vector Machine, Perceptron, and Logistic Regression algorithms incrementally. However, similar to Heusinger et al. (2020a), the process is not fully incremental because it uses TF-IDF and principal component analysis (PCA) for dimensionality reduction.

Abid et al. (2018, 2019) presented a method for text stream clustering called AIS-Clus, based on the artificial immune system (Kephart et al., 1994). This system has online and offline phases. The offline phase comprises receiving historical data to generate the first clusters. In the online phase, new data are divided into equal blocks, i.e., it works in batches. Concurrently, each instance is evaluated alone, being also capable of handling novel classes. Thus, this work could be categorized as Incremental in Batches or Incremental with One input at a time, depending on the point-of-view. Although it works in a clustering fashion, the method performs classification tasks.

Murena et al. (2018) presented the adaptive window-based incremental LDA (AWILDA), a method for topic modeling in document streams. This method contains two LDA models, one for topic modeling and another for drift detection, with the help of ADWIN. It receives the data in batches, making it possible for the approach to use ADWIN as a drift detector and to resort to LDA over the batch.

Assenmacher and Trautmann (2022) presented a stream text clustering method. The use of online and offline phases for algorithms that perform stream clustering is well known. The offline phase generally performs adjustments in the model, such as the stale cluster removal and merging of similar clusters. In the online phase, the method receives input data and verifies the most similar cluster to assign the new input data to the most similar cluster. However, a new cluster is created to accommodate the incoming text if no cluster is sufficiently similar. Due to these characteristics, this method could be categorized as Incremental with One input at a time. Interestingly, this method outperformed other batch-based methods in the evaluation considered in the paper.

The single representative of this model update category is Amba Hombaiah et al. (2021). As aforementioned, this study proposes three methods for sampling to update the language models. The three methods, i.e., the Token Embedding Shift method, Sentence Embedding Shift method, and Token MLM Loss method, use both current and previous models to evaluate changes to sample new data to fine-tune the current model. Thus, more significant differences between a given text representation and the representations provided by the old and current models generate higher chances for a given text to be selected for fine-tuning. Thus, in this specific case, it is costly to fine-tune using all the data because of the size of the BERT models. In addition, GPUs are necessary to speed up the training/update of these models.

Some papers resort to the complete retraining of models. The retraining can be triggered by drift detection or periodically, typically after batch processing. As noted, Chamby-Diaz et al. (2019) proposed a dynamic feature selection method to handle feature drift, namely Dynamic Correlation-based Feature Selection (DCFS). This method uses concept drift detectors, such as ADWIN. Concept drift detectors generally provide two levels of signaling: warning and drift. Whenever a warning signal is outputted, DCFS updates the covariance matrix incrementally. The feature-feature and feature-class correlations are calculated when a drift signal is emitted. Thus, a new Naive Bayes model is trained from scratch using the feature subset selected according to the correlation-based feature selection (CFS).

Other works such as D’Andrea et al. (2019), Bechini et al. (2021), and Bondielli et al. (2022) also utilize the retraining scheme. All these papers compare approaches that resort to the retraining scheme. Retraining occurs regularly and considers data from events. However, the dataset is increased incrementally to be used by the methods during the training step. For example, when event #10 concludes, the data related to this event are appended to the data regarding previous events. Thus, a new model can be trained based on the dataset, now containing the data about event #10. Concerning these three works, only Bondielli et al. (2022) uses an incremental approach, i.e., Complement Naive Bayes (Rennie et al., 2003) with the partial fit. For this approach, however, the authors used TF-IDF for vectorization, which is not updated during the online monitoring after the first event. Thus, the process is not fully incremental. In addition, the authors did not mention any strategy for maintaining a dataset in a feasible size after several incremental additions of batches.

The system proposed by Susi and Shanthi (2023), i.e., TSDA-BERT, also considers periodic retraining to overcome sentiment drift. Whenever a sentiment drift happens, the system uses a domain impact score, which calculates the impact of a tweet in the domain. The calculation considers the intersection of a tweet’s words and the domain-specific impact words. According to the authors, if the impact is above 0.5, it indicates adherence to the domain. However, the authors do not explain how the domain-specific words are selected. Compared to D’Andrea et al. (2019); Bechini et al. (2021) and Bondielli et al. (2022), Susi and Shanthi (2023) provided a strategy to maintain the training set in a feasible size. The tweets with higher adherence to the domain are included in the training set, and the same number of tweets are removed from the training set. It means that the training set is always the same size. The authors mentioned the utilization of at most 324,685 tweets in the training set. This training set is used for fine-tuning over time.

Classification is among the most common stream mining tasks. In the general classification, the objective is to predict, with arbitrary accuracy, a unique class from a small set of values from a given input. Some applications found in the papers addressing the classification task include (i) crisis management; (ii) fake news detection; (iii) fake review detection; (iv) hashtag prediction; (v) sentiment analysis; (vi) short-text classification; and (vii) spam detection.

Regarding crisis management, Pohl et al. (2018) aimed to identify the relevant tweets about two environmental disasters: the Colorado floods and the Australian bushfires. It is considered a binary classification task because the model assesses whether or not a tweet is relevant, sometimes with a human in the loop. Their approach is evaluated regarding the average error and the number of queries. Because the method presented in Pohl et al. (2018) employs active learning strategies, the label uncertainty determines whether the system should query a user. Only Pohl et al. (2018) represents this application in the classification task.

Fake news detection is addressed in Fenza et al. (2023). The authors proposed an index based on fuzzy formal concept analysis, which correlates with the classifier’s performance. According to the authors, the fake news detection problem is generally tackled as a binary classification, where a model should classify news as fake or real. The authors evaluated three machine learning methods: Randon Forest, Naive Bayes, and Passive-Aggressive (Crammer et al., 2006). Although the authors proposed the method, they did not couple the index to the methods to trigger retraining. Three datasets containing news articles between 2018 and 2020 were used, i.e., NELA-GT-2018, NELA-GT-2019, and NELA-GT-2020 (see Section 5). According to the authors, only the Passive-Agressive algorithm was tested online, and the news between February and August 2018 were used as the training set, considering also the fuzzy lattice structure. The classifiers’ evaluation is performed using accuracy and F1-score. The evaluation of the proposed index happens through visual analysis, Pearson’s and Spearman’s correlation, and cosine similarity. The authors argue that their method would allow early drift detection but do not provide experiments or evidence.

Considering fake review detection, Mohawesh et al. (2021) tackled this task using three ML methods: SVM, logistic regression, and perceptron. The authors used four Yelp datasets, only one containing fake and genuine reviews. Mohawesh et al. (2021) noted that the datasets “were built based on an unknown filtering algorithm and web-scraper techniques to label each review as fake or genuine”. Because the idea is to determine whether or not a review is fake, it corresponds to a binary classification task. The ML methods are evaluated using accuracy and statistically assessed using the Nemenyi test (Nemenyi, 1963). The authors claimed that their work is the first to address concept drift in the fake review detection problem. Considering the selected papers, this is the only method that approaches fake review detection.

Hashtag prediction is addressed in Heusinger et al. (2020a), Amba Hombaiah et al. (2021), and Heusinger et al. (2022). Both Heusinger et al. (2020a) and Heusinger et al. (2022) used random projection as a dimensionality reduction method for text streams. Also, a dataset, i.e., NSDQ, is proposed for the problem because it generates high-dimensional data and can be reduced in real-time by random projection. Furthermore, this is the only real-world textual dataset addressed in these papers, while the others are synthetic. This dataset contains 15 classes that make the stream mining task approached by them a multiclass classification. The evaluation is performed in terms of accuracy, Cohen’s Kappa, and run time. In Amba Hombaiah et al. (2021), the sampling approaches for updating BERT are tested using two datasets: OffensEval 2019 and Country Hashtag Prediction. Approaching OffensEval constitutes a binary classification task; thus, the authors used the Area Under Curve (AUC) of the Receiving Operating Characteristic (ROC) curve and F1 score. However, addressing Country Hashtag Prediction corresponds to a multiclass classification task and is evaluated using micro-F1 score, macro-F1 score, and accuracy.

In sentiment analysis, the objective is to develop a model capable of inferring a user sentiment from text. According to Medhat et al. (2014), “sentiment analysis (SA) or opinion mining (OM) is the computational study of people’s opinions, attitudes and emotions toward an entity”. Similarly to sentiment analysis, stance detection regards the position of a given text’s author about a specific topic, considering the labels in favor, neutral/neither, and against, sometimes expressed in literature with different labels but with similar meanings (Bechini et al., 2021; Küçük and Can, 2020). The papers D’Andrea et al. (2019), Bechini et al. (2021), and Bondielli et al. (2022) approach stance detection, with D’Andrea et al. (2019) and Bechini et al. (2021) related to vaccination, and Bondielli et al. (2022) regards the stance about the green pass, as mentioned in previous sections. The authors in these three works collected the dataset that they needed to utilize primarily from Twitter. As aforementioned, stance detection classifies texts in three labels, indicating that it is a multiclass classification task. In D’Andrea et al. (2019), the metrics utilized to evaluate the method are F1 score, precision, recall, AUC, and accuracy. In Bechini et al. (2021), models were evaluated using accuracy and F1 score, and Bondielli et al. (2022) used F1 score, accuracy, and the number of features in each model.

Bravo-Marquez et al. (2022) proposed a sentiment lexicon inductor for time-evolving environments in a sentiment analysis context. The authors claimed that sentiments could change over time, while new words in different sentiments can emerge. In addition, the lexicon would be static in a fully incremental system without sentiment induction. In this case, from a seed lexicon, the authors processed the dataset in a stream fashion and, at the same time, inferred sentiment from tokens absent in the lexicon. Although, in practice, the system outputs a value limited by a logistic function, we presented this paper in the classification section because of the sentiment analysis application. In addition, the authors tested their approach by deliberately changing lexicon sentiment scores and measuring how long the system would take to recognize the new sentiments. Finally, the authors used accuracy and Cohen’s Kappa to evaluate the classifiers applied together with their method.

Short-text classification is addressed in papers such as Li et al. (2018); Sun et al. (2021) and Li et al. (2022). Li et al. (2018) proposed a method for short text streams bound to concept drift. This method takes advantage of Probase for short text enrichment. The approach is evaluated in terms of time and accuracy. Sun et al. (2021) described a method for text stream classification based on feature extension and ensembles formed by ensembles. This method can handle concept drifts by calculating the distance between each short text in the previous and new batches. Li et al. (2022) proposed a method for short text classification in text stream scenarios. This method enriches text by using representations from BERT and Word2Vec. In addition, the method uses a Convolutional Neural Network (CNN) to extract high-level features. This method handles concept drift by resorting to a concept drift factor used in the systems. Both approaches in Li et al. (2018); Sun et al. (2021) and Li et al. (2022) are applied to the same datasets, i.e., Tweets, TagMyNews, and Snippets, considering text classification as topics.

Some papers address Spam detection as experiments (Melidis et al., 2018; Chamby-Diaz et al., 2019; de Moraes and Gradvohl, 2021). Because the goal is to classify a piece of text into either non-spam or spam, the task is considered a binary classification task. The authors in Melidis et al. (2018) provided an ensemble-based mechanism for predicting a feature’s probability of association with a given class by considering that words might be subject to temporal trends and a sketch-based feature space maintenance mechanism that allows for memory-bounded feature space maintenance. The approach utilizes an ensemble compounded by statistical techniques to account for feature periodicities. The ensemble consists of a Poisson model (Melidis et al., 2018), a Seasonal Poisson model (Holt, 2004), an Auto-regressive Integrated Moving Average (ARIMA) model (Box et al., 2015), and an Exponential Weighted Moving Average (EWMA) model (Nishida et al., 2012), to capture regular, seasonal, auto-correlated, and sudden trends. A sketch-based approach is designed to maintain a concise feature space. The authors tested three versions: a baseline sketch that retains only word and occurrence counts, a fading sketch that considers the importance of frequent words, and a drift-detector-based sketch, which uses ADWIN to detect the decrease in word usage. The approaches are compared regarding the accuracy, Cohen’s Kappa, and run time. Chamby-Diaz et al. (2019) proposed a method for feature selection based on correlations to handle feature drifts in data stream scenarios. The method is not exclusive to spam detection, but the spam dataset is the only text-based dataset used by the authors. The method is evaluated in terms of accuracy. de Moraes and Gradvohl (2021) proposed a method for feature selection in binary text stream classification tasks, namely OFSER. The proposed method leverages adaptive regularization and weighs the input for each new data. The regularization, according to the authors, decreases the impact of the feature drift. Despite being fast and having decent overall performance, their method depends on a parameter to define the number of features to be selected from the original set. The method runs on top of a naive Bayes classifier, chosen due to its simplicity and naive assumption of independence among the features. The approach is evaluated using F1 score, accuracy, memory consumption, and run time. Furthermore, due to “an undesired conservativeness of the Friedman test” (de Moraes and Gradvohl, 2021), it is statistically assessed using the Iman-Davenport test (Iman and Davenport, 1980) instead of the Friedman test (Friedman, 1937, 1940), and the Bergmann-Hommels’ procedure (Garcia and Herrera, 2008) instead of the Nemenyi test (Nemenyi, 1963). OFSER ranked among the three best approaches.

As expected, the most frequent metrics in this stream mining task are: accuracy, Cohen’s Kappa, F1 score, AUC, and run time. Although not all methods were assessed regarding run time, it is crucial to have values for this metric due to its use in streaming scenarios, where time and memory consumption are constrained.

Clustering is a stream mining task in which the aim is to find intrinsic clusters, according to their features (Bifet et al., 2018). The general idea is to minimize the similarity between different clusters and maximize the intra-cluster similarity (Bezerra et al., 2015). Differently from classification, in the clustering task, the labels are not available before the learning process. Therefore, alternative metrics are necessary and, since there is no ground truth, the learning process is named unsupervised (Bifet et al., 2018).

Three works approach the stream clustering task (Abid et al., 2018, 2019; Assenmacher and Trautmann, 2022). The first two papers presented similar approaches that use the artificial immune system (AIS) for text clustering, while the third presents textClust, a stream clustering method. Abid et al. (2018) developed a method for text stream clustering based on the AIS called AIS-Clus. It uses heuristics based on the AIS to cluster data efficiently and, by discovering these clusters, can also detect concept drift and feature evolution. The authors could also recognize new classes corresponding to the concept evolution task in the experiments. According to the authors, the AIS is analogous to the biological immune system because it receives an intruder, clones specific cells, and handles the intruder until it dies. In their approach, for each new input (analogized as antigen), a scoring function calculates its adherence to each cluster (analogized as a B-cell). The clonal selection makes copies of clusters that undergo a mutation process. Later, the negative selection mechanism makes it possible to detect noisy data. Their method does not start from scratch, having an initial static phase for preprocessed historical data clustering. The other phase is online stream processing, which receives the clusters from the first phase as input. The authors used a survival factor for each word in an aging-like scheme. Although it works in a clustering scheme, the method is evaluated regarding F1 score, accuracy, recall, and precision. More information is provided in Abid et al. (2019), which expands on Abid et al. (2018), and new experiments are executed. For example, to test the approach’s capacity to handle new classes, the authors arranged data in three datasets to simulate the emergence of new classes/events, one of which included texts in Arabic. When AIS-Clus is compared to other methods, i.e., CluStream and DenStream, it achieves the best results regarding the precision, recall, and number of clusters, functioning as a classifier as described in Abid et al. (2018).

Assenmacher and Trautmann (2022) presented an online method for textual clustering, namely textClust. The algorithm is available within RiverML Python library (Montiel et al., 2021)⁹⁹9https://riverml.xyz/0.19.0/api/cluster/TextClust/. Over time, in the offline phase, the model is maintained concisely by merging similar clusters and removing outdated ones. A fading factor for the cluster weighting is used to determine cluster staleness. The method was evaluated regarding homogeneity, completeness, and Normalized Mutual Information. Homogeneity evaluates how well a clustering method assigns the data points to the clusters. Reaching 1 for homogeneity means that each cluster contains data points of a single class. On the other hand, completeness measures whether the data points of a given class were assigned to the same cluster. Reaching the value 1 for completeness means that the data points of each class were assigned to a single cluster. The authors support these statements by mentioning that “completeness scores tend to be lower than the homogeneity scores”, and that it “indicates that online clusters are quite pure with low entropy, but the topics are distributed over multiple clusters” (Assenmacher and Trautmann, 2022).

Most selected works that addressed a stream clustering task focused on short-text clustering. Rakib et al. (2021) proposed an efficient method for similarity-based short-text stream clustering called EStream. The method’s efficiency comes from utilizing an inverted index to find the most similar clusters. The authors tested lexical (unigram, bigram, and biterm) and semantic text representations (using a pre-trained GloVe (Pennington et al., 2014)). Their method has two steps: the online and the offline phases. First, each cluster is lexically represented as a cluster feature 4-sized vector consisting of the features (in unigram, bigram or biterm), their frequencies in the cluster, the number of texts in the cluster, and the cluster identifier. The semantic representation consists of the cluster vector and the cluster center, calculated from the average of the GloVe representation of the texts. EStream is compared in terms of Normalized Mutual Information (NMI), Homogeneity, and V-Measure. EStream had the best performance in 50% of the datasets used for evaluation. The authors highlighted that EStream requires less running time and that it stores more information than the other approaches, but that would be an acceptable trade-off (Rakib et al., 2021). They also highlighted that EStream might perform inadequately in more extensive texts.

Vo (2022) proposed a new method called GOWSeqStream, for short text stream clustering, using deep sequential methods, graph-of-words representation, and pre-trained word-embedding models. It uses subgraph mining to extract semantic information from the texts, although it lacks information on how to use it, considering even the number of sliding windows and the support. The method also utilizes Word2Vec representations to generate embeddings to serve as input for other deep encoders, such as GRU. The author also experimented using bidirectional LSTM, Doc2Vec, and BERT representations. These representations are utilized as input for a Dirichlet Process Mixture Model (DPMM). The method was compared with five approaches using three datasets; the proposed approach achieved the best values for two. The author also compared the representation generation; the best combination was with BERT and Bi-LSTM. In addition to English, the author used a Vietnamese text dataset as a final test. In this scenario, the proposed approach achieved the best results among the competitors. As in Rakib et al. (2021), the authors used the Normalized Mutual Information (NMI) as the primary evaluation metric.

Yang et al. (2021) proposed a new short text stream clustering method using an incremental word relation network. The authors highlighted their primary contribution as (a) a new method for real-time short text clustering using a bi-weighted relation: term frequency and co-occurrences, to overcome sparsity; (b) a fast method to locate core terms that represent text clusters sufficiently; (c) the mechanism to overcome topic drift, removing outdated relations and incrementally adding new terms and relations. In addition, the authors proposed a new data structure to represent the clusters, which they named cluster abstract. This data structure has five fields: an index, the number of short texts in clusters, the sum of timestamps, the squared timestamps sums, and a new attribute compared to EWNStream (their previous approach) called pd, containing a core term set. The method uses data windows and specific calculations to update the model to add new data, exclude outdated data, and merge clusters. Besides, the method has a decay scheme to control the forgetfulness of old clusters. In essence, the method develops a graph containing terms and relations, and the clusters are obtained from groups of closely related words. The method searches for a cluster abstract with the most intersection of words considering the input data to predict a cluster to newly inputted data. Using a dataset crawled by themselves, the authors compared their proposed method against EWNStream, MStream, Sumblr, and Dynamic Topic Model. EWNStream+ outperformed its previous version (achieving roughly 86% of NMI accuracy) and was approximately 30 percentage points better than MStream, the third in the ranking. In addition, the run time was very modest across different stream lengths.

Yin et al. (2018) proposed two text stream clustering algorithms: (a) MStream, a one-pass clustering method that utilizes Dirichlet Multinomial Mixture Model (DPMM) and an update process per batch; and (b) MStreamF, which deletes outdated clusters, maintaining a concise model. Considering the clustering process of the MStream algorithm, there is the assumption that the new documents arrive sequentially, and each is processed only once. The initial document generates a new cluster, and subsequent documents choose one of the existing clusters or create a new one. The authors’ updating process proves beneficial in the batch processing of text streams. The process is designed such that each document gets assigned and then temporarily deleted from the cluster so that the similarity of the other document in the same batch is not impacted. After completing the batch process, all documents are assigned to their original cluster. For MStreamF, the authors developed a deleting scheme that works for batch processing by adding a new parameter $B_{s}$, which accounts for the number of batches. When the number of processed batches meets the $B_{s}$ parameter, the new batches are processed after the documents related to the oldest batch are deleted. As the iterations go by, it is expected that some clusters become empty, indicating that they are outdated and could be deleted. The approaches were assessed in terms of NMI, run time, and the number of clusters. They concluded that MStreamF is faster than MStream due to the conciseness of the former model. Comparing the proposed and the state-of-the-art models, MStream and MStreamF outperformed their competitors. MStreamF performed best with temporally ordered datasets, whereas MStream performed best with unordered datasets. The run time of all algorithms increased linearly with the size of the datasets, while the single-pass algorithms were faster.

In summary, the NMI, run time, and the number of clusters are the most often used metrics for stream clustering and short-text stream clustering. The latter may be considered a measure of conciseness, which directly corresponds to one of the constraints of streaming scenarios, i.e., memory consumption, and may indirectly impact run time. NMI, a Shannon-entropy-based metric, measures the similarity of two sets and, concerning clustering, the similarity of the ground-truth and the model-generated clusters (Yin et al., 2018; Emmons et al., 2016). Other metrics may appear, such as completeness, and homogeneity. Those metrics vary between 0 and 1, where the higher, the better. Homogeneity evaluates how well a clustering method assigns the data points to the clusters. A perfect homogeneity, i.e., 1, indicates that each cluster contains data points of a single class. As aforementioned, completeness evaluates whether the data points of a given class were assigned to the same cluster. A perfect completeness value suggests that the data points of each class were assigned to a single cluster.

In this category, we grouped papers that tackled event detection and novelty detection. According to Faria et al. (2016), novelty detection is “the ability to identify an unlabeled instance (…) that differs significantly from the known concepts”. As suggested in Faria et al. (2016), we considered concept drift detection, semantic shift detection, and concept evolution as sub-categories of novelty detection.

We also considered physical event detection a sub-category of event detection. As mentioned previously, Suprem and Pu (2019a), Suprem et al. (2019b), and Suprem and Pu (2019b) described distinct aspects of a system for landslide detection. They utilized governmental reports as trustworthy sources and social media posts as social sensors (also named strong and weak signals, respectively). The system is described as fully autonomous and continuously evolving, becoming unnecessary human intervention. Although the works are similar in several aspects, there are minor variations in the evaluation metrics. In Suprem and Pu (2019a), the selected metrics are precision and F1 score. The event detection is assessed using false positives and false negatives, where the original variant of the system is used as ground truth. Suprem et al. (2019b) used F1 score, precision, recall, and the number of events detected as metrics. There is no ground truth regarding the number of events: only the events counted. Suprem and Pu (2019b) used accuracy to evaluate classifiers’ performance across data windows.

Kolajo et al. (2022) proposed a framework for real-time event detection using social media as a data source. The interesting highlights in this paper regard the tweets’ enrichment for slang, abbreviations, and acronyms based on external sources. The method creates a local vocabulary using data from various external sources. In addition, they utilized spelling correction and emoticon replacement. The authors used an incremental clustering algorithm to cluster events and then rank these events based on important words for each event. The authors evaluated their method using two experiments: (a) comparing it to the General Social Media Feed Preprocessing Method (GSMFPM) to determine if the enrichment layer performs effectively; and (b) event detection from social media. In experiment (a), the authors represented the tweets using unigrams and bigrams, supposedly later converted to GloVe (unclear in the paper). Later, the vectors are applied as input to a Feedforward Neural Network (FNN) and a Convolutional Neural Network (CNN). These approaches are not incremental, thus presenting concerns about the process’ timeliness regarding real-time events. In this experiment, they measured the cross-entropy loss across the training epochs for both Twitter Sentiment Analysis and Naija datasets. Their method outperforms GSMFPM. The second experiment measures accuracy over events in social media, using precision, recall, and F1 score. The authors used a dataset called Event2012, which contains annotations about events. The proposed method obtained a higher F1 score than the other approaches.

Regarding novelty detection and its subdivision in this study, only one paper exclusively considers concept drift detection (de Mello et al., 2018). Three include the concept evolution problem (Abid et al., 2019; Chamby-Diaz et al., 2019; Wang et al., 2019), and another mentions the semantic shift detection (Periti et al., 2022). Considering the concept drift detection, de Mello et al. (2018) used a cross-recurrence quantification analysis (CRQA) to detect concept drifts. The author’s idea was to highlight the most significant hashtag-related events. Cross-recurrence quantification analysis is used to compare the changes in trajectory. This outcome is achieved by assessing the longest diagonal line of two consecutive windows and whether they follow the same generating process over time. All operations occur inside a system called TSViz. The experiments discussed in the paper are on drift detection related to hashtags from Brazilian politics. The authors concluded that the drifts detected directly trace back facts from the news. According to the authors, recurrence analysis “characterizes the behavior of dynamical systems by reconstructing produced data in phase spaces”. The authors used Normalized Compression Distance (NCD) to compute the similarity among texts and Naive Bayes to perform sentiment analysis; however, the authors did not detail the classification process. The results were visually assessed.

The concept evolution problem regards the increase in the number of classes over time. For a model to be updated, it must internally account for these novel classes (Faria et al., 2016). Traditional ML methods require prior knowledge of the number of classes. Abid et al. (2018) and Abid et al. (2019) proposed a method for text stream clustering based on AIS. These papers were previously referenced in this work. They also managed concept evolution (under the name of novelty detection). These methods address the concept evolution problem by cloning and mutating existing clusters, a heuristic of the clonal selection principle. If the novel data do not fit into a cluster, they are sent to the outlier buffer, where they are examined periodically to detect novel classes. Abid et al. (2018) evaluated the quality of concept evolution handling using the $M_{new}$ metric, which measures the rate of novel class instances misclassified as from an existing class. In addition, the authors plotted the F1 score, accuracy, and recall over time, demonstrating the appearances of new classes and how their method recovers from concept evolution. The run time was not measured. Abid et al. (2019) employed a similar plot as Abid et al. (2018) for two datasets. In addition, they plotted the number of existing classes and identified classes by the method over time. The metric $M_{new}$ is also used, and the number of missed classes is computed.

Wang et al. (2019) proposed ESACOD, a framework for streaming classification with concept evolution and subject to concept drift. Their work aims to learn satisfying parametric Mahalanobis-based metrics in real-time. According to the authors, the objective is to identify a feature space projection in which its constraints generate properties of cohesion and separation (Wang et al., 2019). Cohesion is the ability of data points to occur close to others from the same class. In contrast, separation is the ability of data points to be distant from others from different classes (Wang et al., 2019). Their method trains an open-world classifier with a small dataset with an initial metric established. When new data arrives from the stream, the metric is applied to it, generating data in a new feature space, and the prediction is made afterward. If the prediction indicates that the data does not belong to a novel class, the prediction remains unchanged. On the contrary, if the classifier assumes the data are from a potentially novel class, the data are added to a buffer. When this buffer is filled, it is checked for concept evolution and concept drift. An arbitrary percentage (between 0 and 30%) of data with their respective labels is required. Finally, the evolution class metric is computed using paired constraints based on this randomly selected data. Later, a K-Means algorithm (Lloyd, 1982) is applied, and a label propagation (Zhu and Ghahramani, 2002) method is performed apparently to the other data in the buffer. If a concept drift or concept evolution is detected, a new classifier is trained with the data to replace the older classifier. The authors concluded that their approach could address the challenges of multiple novel class detection and stream classification bound to concept drift and with few labels available. The method is evaluated in terms of accuracy and run time. Concerning concept evolution, the metrics used are $M_{new}$ and $F_{new}$, which measure the instances of an existing class misclassified as a novel class, $A_{new}$, which is the accuracy of novel class classification, and $A_{known}$, which is the accuracy of known class classification.

Regarding semantic shift detection, Periti et al. (2022) addressed this problem in an incremental way. The authors used incremental clustering techniques (such as affinity propagation) to generate representation clusters in time slices. The word contexts in the past are clustered into several clusters, serving as a memory for posterior observations. To generate representations, the authors tested BERT and Doc2Vec. BERT provides contextual representation, whereas Doc2Vec provides pseudo-contextual embeddings. The approach selects documents in which target words emerge, fine-tunes the embedding model to add new arriving documents, extracts the embeddings, clusters the representations, and refines the clusters by removing clusters of single or old representations. The authors tested their approach using representations generated by BERT and Doc2Vec, for two datasets from SemEval 2020: CCOHA and LatinISE. The authors evaluated alternatives based on affinity propagation. The incremental version of the affinity propagation (IAPNA) performed adequately on the LatinISE dataset using BERT representations and on the English dataset using Doc2Vec representations. In contrast, the affinity propagation a posteriori had satisfying results in the opposite situations. The authors were surprised that Doc2Vec obtained decent results and consumed less time than contextual models. It is important to notice that the target word must be known in advance to perform analyses.

Topic modeling consists of statistical tools to examine textual data and identify the most relevant terms related to each theme. This approach facilitates the exploration of the interconnections among these themes and their temporal evolution (Blei, 2012). It is also considered a text mining task (Kherwa and Bansal, 2019). Four selected papers approach topic modeling (Murena et al., 2018; Hu et al., 2018; Van Linh et al., 2022; Nguyen et al., 2022).

Murena et al. (2018) proposed an approach mixing LDA and ADWIN to overcome the problem of topic modeling in document streams, entitled AWILDA. LDA (Blei et al., 2003) is a common method for topic modeling. The authors mentioned that LDA had gained much attention, and it also has an online version. However, one problem with the online version is setting window sizes because drifts may happen in a smaller period than the window size. Thus, the authors defined the window with the aid of an ADWIN module, which can assist in determining topic drifts and the new window for LDA to consider. Two classes of algorithms are mentioned: the passive, which updates a model for each observation, and the active algorithms, which attempt to detect the drift and update the model only when the drift is detected. We can draw parallels between these classes of algorithms and the detection methods presented in Section 4.3, i.e., adaptive and explicit, respectively. The author’s idea was to separate the task of topic modeling and topic drift detection. There are two LDA models inside AWILDA: one for language modeling ($LDA_{m}$) and the other for drift detection ($LDA_{d}$). In this approach, for each document received from the stream, AWILDA reckons the likelihood for $LDA_{d}$ and adds it to the ADWIN module. If a drift is detected, $LDA_{m}$ is trained on the subwindow ADWIN selects. $LDA_{m}$ is updated whenever a new document arrives from the stream. The authors evaluated their proposed method using the perplexity metric for document modeling and the latency between the actual current drift and the detection. According to the authors, perplexity is “used by default in language modeling to measure the generalization capacity of a model on new data” (Murena et al., 2018). The authors concluded that AWILDA could recognize all drifts in the synthetic datasets and one version of the real-world dataset. In addition, the method can select the documents window to be used for updating. AWILDA can detect abrupt drifts and works sufficiently for gradual drifts. Compared to online LDA, it works similarly until a drift occurs. When drift occurs, AWILDA is retrained, which increases perplexity, but it outperforms the online LDA ultimately.

Hu et al. (2018), proposed a short text stream classification method that uses content expansion and includes a concept drift detector. According to the paper, the external sources must satisfy two criteria: to be large and sufficiently rich to comprise most contents in the short text stream that will be classified and highly topic-consistent with the text stream. The method mines hidden information from the external corpus by using LDA because, according to the authors, LDA performs adequately on longer texts. From the LDA model, top representative words for the topics are selected to be added (once or several) times to a short text according to the topic distribution and word probability of belonging to a topic. The topic distribution represents each short text. The method was evaluated regarding accuracy (classification task) and the drifts, using false alarms, missing drifts, and delays. The datasets were arranged to simulate drift; however, the method is unspecified. The authors concluded that their approach surpasses the accuracy of all the competitors, demonstrating more stability. In addition, their approach could recover from drift earlier than other approaches and outperforms the competitors regarding delay and missing drifts.

Van Linh et al. (2022) proposed a graph convolutional method for topic modeling, considering short and noisy text streams. The authors leveraged Word2Vec representations and Wordnet knowledge graph to improve the predictions of their method, called GCTM. The authors claimed that their method could balance the knowledge graph and the knowledge obtained from the previous data batch. This ability can be valuable when handling concept drift. GCTM integrates a graph convolutional network (GCN) into an LDA model to exploit a knowledge graph, and both are updated simultaneously in the streaming environment. The authors tested their approach using six short text datasets and two regular text datasets. Using previous knowledge allows GCTM to output satisfying predictions and recover more quickly from concept drift. The authors simulated concept drift by rearranging the topics sequentially. The metrics selected for evaluation were the Log Predictive Probability (LPP) (Hoffman et al., 2013) and the Normalized Pointwise Mutual Information (NPMI) (Lau et al., 2014). These methods measure the model generalization and the coherence of the topics, respectively. GCTM is evaluated in two ways: utilizing Word2Vec (GCTM-W2V) and the knowledge from the Wordnet graph (GCTM-WN). GCTM-WN and GCTM-W2V outperform the competitors in LPP across all the datasets, even in the presence of concept drift. The authors also performed an ablation study.

Nguyen et al. (2022) proposed an LDA-based topic modeling approach with mechanisms for balancing stability and plasticity, namely BSP. Stability-plasticity is a dilemma involving maintaining old knowledge (stability) and learning new knowledge (plasticity) (Nguyen et al., 2022; Gama et al., 2014). Balancing both prevents concept drift from impacting performance and catastrophic forgetting (Nguyen et al., 2022). The authors used TPS and iDropout combined into an LDA-based topic modeling method. TPS (Tran et al., 2021) aids the model with external knowledge, i.e., Word2Vec representations. iDropout (Nguyen et al., 2019) created variables $\beta^{t}$, updated whenever a new mini-batch is inputted. Because both are different mechanisms, the authors modified the calculation of $\beta$ to comprise information from both mechanisms. They performed experiments on eight datasets: one long, two regular, and five short-text. The authors compared their method to six different approaches. The hyperparameters were selected using a grid search. Similar to Van Linh et al. (2022), the authors contrasted LPP and NPMI. The authors tested using the datasets shuffled and ordered chronologically whenever possible. Their method achieved the best values for LPP in four out of six datasets tested. It is worth noting that their method achieves satisfactory results very rapidly at the highest levels. Their method maintains high levels of performance while using chronological datasets. The authors tested the stability and plasticity by simulating drifts by sorting the topics in order of classes, similarly to Van Linh et al. (2022). BSP could reach the best values when testing for catastrophic forgetting and maintained the highest levels when recovering from concept drift. As in Van Linh et al. (2022), the authors performed an ablation study to understand the impact of some parameters.

Bigram adheres to the Bag-of-words concept, in which it is possible to organize texts in two dimensions: columns as words, and rows corresponding to documents. The cells contain the count of a given word in a specific document. The difference is that a pair of sequential words is represented in each column instead of the words. For example, the sentence “he has been here” will generate three columns: (he, has), (has, been), and (been, here). The challenge incurred from utilizing bag-of-words in streaming scenarios also happens to bigrams, i.e., the dimensions regard fixed words and do not evolve. Rakib et al. (2021) used three representation methods while testing their proposed method for short-text stream clustering: unigram, i.e., bag-of-words, bigram, and biterm. Assenmacher and Trautmann (2022) also used both unigram and bigram for text representations.

According to Hu et al. (2018), a biterm corresponds to unordered word-pair co-occurrences. Furthermore, the authors in Hu et al. (2018) highlighted that biterms are more sparse than regular bag-of-words, and utilized external sources to reduce the sparseness. Considering the biterm definition and using the same example as in a bigram, the biterms generated from the sentence “he has been here” would be (he, has), (he, been), (he, here), (has, been), (has, here), and (been, here). Considering the text stream scenario, it encounters similar challenges as bag-of-words and bigrams. To overcome this, Hu et al. (2018) developed an ensemble based on base learners trained using data chunks, each with its biterm topic model. Rakib et al. (2021) also used biterm as text representations. To evaluate their short-text stream clustering method, the authors used biterm, unigram, and bigrams. Biterms performed better than bigrams and unigrams, considering NMI values.

Co-occurrences count simultaneous occurrences of two particular words. Yang et al. (2021) developed a bi-weighted word relation network that considers both the co-occurrences and the word frequencies. Although co-occurrences and word frequencies are not representations, we opted to include them as a single representation because they will be part of a graph, i.e., graph-of-words, which is an actual representation.

Graph-of-words (GOW) is a textual representation that transforms documents into graph-based structures (Vo, 2022). It can maintain long-term relationships between words, according to the author. After generating the graphs regarding specific documents, frequent subgraph mining techniques are applied, and later the mined frequent subgraphs are used as feature representations. In Vo (2022), GOW has two parameters: sliding window and minimum support. GOW appears to have the capability of being updated in real-time. However, its use with a pre-trained Word2Vec model (that can be outdated after an arbitrary period) makes the process not fully incremental. Although they did not use the terminology graph-of-words, Yang et al. (2021) developed a corpus-level word relation network, namely EWNStream+, which retains the co-occurrence counts and word frequencies. According to the authors, EWNStream+ is incremental by receiving data batches. Van Linh et al. (2022) proposed a a novel graph convolutional topic model (GCTM) based on graph convolutional networks and LDA. The initial graph is formed using words and their relations. GCTM is tested using Word2Vec representations and WordNet. GCTM does not support incremental-fashion training, implying that the text models can become obsolete.

Word frequency is the word count. Yang et al. (2021) included word frequency as part of their word relation network, which also considers the word co-occurrences. This representation is also used to determine whether a word is outdated in the graph representation.

Several papers chose to use the words themselves rather than any text representation. Murena et al. (2018) presented AWILDA, an LDA-based method integrated with ADWIN for topic drift detection. The authors used the words lowercased and stemmed. Abid et al. (2018) and Abid et al. (2019) described AIS-Clus, an incremental clustering method. Initially, the authors used DBSCAN (Ester et al., 1996) to generate the cluster, and then sketches were developed to summarize each cluster. The sketches contain lists of words and outliers present in a cluster. Hammer and Yazidi (2018) presented a method to handle concept drift in an abruptly changing environment. The authors used words to monitor probabilities in topics. Considering the updating scheme, words could easily be added or removed from sketches. Therefore, we consider it possible to use it in streaming scenarios, although it can become complex and time-consuming to maintain a list of words in every sketch, as demonstrated in Murena et al. (2018) and Abid et al. (2019), if not limited to respecting the constraints of data stream environments.

Bag-of-words (Harris, 1954) is probably one of the simplest methods for vectorizing texts, as it divides the text into tokens or words. Considering rows and columns, these tokens function as columns while the rows represent each text, such as tweets. There will be the counts of the tokens corresponding to a particular column in a text corresponding to a specified row in each cell. An evident characteristic is that bag-of-words representation in a unigram way does not represent the order of words, which can be leveraged in some applications. In streaming scenarios, it inhibits ML methods from performing properly. For example, suppose a bag-of-words representation is generated whenever each new text is inputted. In that case, the number of columns may increase, and most ML methods cannot handle dimension-changing inputs. Furthermore, even if the process runs in batches, the words of the bag-of-words may change. If the first batch defines the words to use in the bag-of-words representation, it may not recognize changes and new words, i.e., new dimensions, over time.

Term-Frequency-Inverse Document Frequency (TF-IDF) is a statistic from the information retrieval area used for determining the importance of words to a document or a set of documents (Salton and Buckley, 1988). The calculation considers the frequency of a term, and the inverse document frequency, which defines how informative a term is across several documents. Generally, TF-IDF is used, in the stream setting, to encode data batches. It is worth noting that the term frequency calculation is remarkably similar to the bag-of-words procedure. Thus, it is common to discover the use of bag-of-words with TF-IDF. Pohl et al. (2018), Bechini et al. (2021), and Bondielli et al. (2022) used TF-IDF after obtaining a data batch to encode the terms and the texts from the stream. Li et al. (2018) utilized TF-IDF to generate the vector representations from the data batches so that a base learner could be trained to be incorporated into the ensemble. Heusinger et al. (2020a) and Heusinger et al. (2022) performed TF-IDF in an offline mode to generate a very high-dimensional vector so that they could test their dimensionality reduction strategy. Mohawesh et al. (2021) executed TF-IDF before all the processing. Later, the authors employed PCA to reduce the dimensionality of the datasets by selecting the 10,000 most meaningful components. Since TF-IDF works together with bag-of-words, it is impossible to update it incrementally because of the change in the dimensions. Assenmacher and Trautmann (2022) used TF-IDF to decide the proximity of incoming text to existing microclusters in the online phase. This calculation is also used in the offline phase, particularly when evaluating the merging of existing clusters. Fenza et al. (2023) used TF-IDF representations to generate the fuzzy lattice structure. Rabiu et al. (2023) leveraged TF-IDF to compute the input vectors to train base learners. An interesting aspect regards preprocessing in Rabiu et al. (2023): the authors utilized the Stanford CoreNLP (Manning et al., 2014) to segment words, part-of-speech tagging, and stemming. The authors used only the first three tags of noun, verb, and adjective. According to the authors, these tags “carry the most valuable information regarding reviewed items”. However, no evidence is provided.

Word2Vec (Mikolov et al., 2013a) corresponds to two distinct model architectures for learning distributed representations: Continuous Bag-of-words (CBoW) and Skip-gram. Both are neural network architectures, where the number of neurons is the same in the input and output layers, and the single hidden layer corresponds to the embedding size. Each neuron in the input and output layers can correlate to the words in the vocabulary. The representations, after training, are often obtained by taking the connection weights between a neuron (representing a word) in the output and the hidden layers. The difference between CBoW and Skip-gram is the training step aim: CBoW aims at predicting a specific word given its surrounding words, whereas Skip-gram does the opposite, i.e., predict the word in the middle based on the surrounding words (Mikolov et al., 2013a). The papers that utilize Word2Vec use it for text representation only. Li et al. (2022) leveraged Word2Vec for reduction of data sparsity. The authors developed their method for short-text classification, and one of the general approaches for this problem is to enrich the data. The authors evaluated both Word2Vec and BERT for the short-text representation, which was later applied to a CNN to extract higher-level feature information. Although Word2Vec is a neural architecture, it has incremental versions by using gensim¹⁰¹⁰10https://radimrehurek.com/gensim/ (Rehurek and Sojka, 2011) or other methods in the literature (Kaji and Kobayashi, 2017; May et al., 2017; Iturra-Bocaz and Bravo-Marquez, 2023).

Le and Mikolov (2014) proposed Doc2Vec to obtain documents as distributional vectors. Doc2Vec is a generalization of Word2Vec. Similarly to Word2Vec, Doc2Vec is constituted by two architectures: Paragraph Vector - Distributed Memory (PV-DM) and Distributed bag-of-words version of Paragraph Vector (PV-DBOW). In PV-DM, the document vectors are trained with the word vectors in the architectures, while in the PV-DBOW, the aim is to predict the words of a document from a document id. Periti et al. (2022) used a Doc2Vec model trained with the CCOHA and LatinISE datasets. The model is not updated during the process and may become obsolete as time passes. In Vo (2022), it is unclear whether the author utilized a pre-trained model, trained a model from scratch, or if the model was updated over time. Since Doc2Vec is a neural architecture, training and updating it can be computationally costly.

Global Vectors (GloVe) is a method for generating co-occurrence-based word vector representations (Pennington et al., 2014). According to the authors, GloVe utilizes global matrix factorization and local context window methods. The method is trained in a batch manner. D’Andrea et al. (2019), Rakib et al. (2021), and Nguyen et al. (2022) used GloVe for semantic representation by using pre-trained models. Kolajo et al. (2022) claimed that GloVe is used for feature extraction. Suprem and Pu (2019a) mentioned that the proposed system, i.e., Adaptive Social Sensor Event Detection (ASSED), supports GloVe. The authors in the original paper (Pennington et al., 2014) did not describe any incremental or adaptive training. Therefore, the vector representations can become outdated over time, constituting a potential disadvantage in streaming scenarios.

FastText, a method presented in Bojanowski et al. (2016) as an extension of the Skip-gram method, is one of the Word2Vec architectures. Instead of accounting for the entire words, FastText considers subword partitions using n-gram vectors. Using an example from the original paper, encoding the word $where$ in a 3-gram fashion results in a 5-sized vector containing (wh, whe, her, ere, re). In addition, the approach incorporates the word $where$ integrally. This method of splitting words in n-grams helps the model handle words unseen in the training step, also named out-of-vocabulary (OOV) words. An incremental update method is not mentioned in the paper. D’Andrea et al. (2019) utilized a pre-trained FastText model (Bojanowski et al., 2016) as a text encoding method. In addition, FastText is used statically, implying that no method is presented in D’Andrea et al. (2019) for the incremental update of the text representations. However, D’Andrea et al. (2019) concatenated FastText representations to bag-of-words representations, generated in each step of an incremental procedure of accumulating data from past events.

Bidirectional Encoder Representation from Transformers (BERT) is a multi-purpose language model that enables several natural language processing tasks (Devlin et al., 2018), such as sentiment analysis, sequence-to-sequence, paraphrasing, and question answering. In addition, BERT can provide vector representations of text to be used in a particular downstream task. Bechini et al. (2021) used an Italian version of the pre-trained BERT model, i.e., AlBERTo (Polignano et al., 2019), for measuring semantic similarity between tweets. In Amba Hombaiah et al. (2021), BERT is the primary model. The authors tested different sampling methods for fine-tuning to pursue an incremental update of the model. BERT is also used as a text encoding method in Vo (2022), where the authors enhanced short text clustering by combining pre-trained BERT’s representations with a BiLSTM and a graph-of-words representation. Periti et al. (2022) used BERT for word representation generation in both English and Latin by using pre-trained models. Considering the aforementioned papers, only Amba Hombaiah et al. (2021) has an updating scheme for the representations. It is achieved by using fine-tuning strategies, which can enable the use of BERT in streaming scenarios while it may also become a bottleneck in the process. Susi and Shanthi (2023) leveraged BERT and variations in two moments. First, the authors used a pre-trained RoBERTa model (Liu et al., 2019) specifically suited for sentiment classification. The RoBERTa model enables automated training data generation. However, another BERT model is fine-tuned in the system whenever a sentiment drift happens. Li et al. (2022) used BERT to enrich short texts. Short texts are very sparse, and, according to the authors, using embeddings may improve the representation quality.

Moghadasi and Zhuang (2020) proposed a sentence embedding method that considers the sentiment score behind the sentence. Kolajo et al. (2022) used the Sent2Vec embeddings to compute the semantic representation of the input texts and then cluster these texts. If a new tweet is different from the histograms of the clusters, a concept drift is deemed to have occurred, and a new cluster is created for it. Kolajo et al. (2022) did not describe an updating scheme. Thus, the Sent2Vec model can become obsolete over time, necessitating retraining.

Bravo-Marquez et al. (2022) proposed a vector representation method for texts that can be considered a table-like representation, with the columns corresponding to words and rows similarly corresponding to words. However, the column (in the original paper, called context) and the words (called vocabulary) can have different sizes. The number of contexts defines the dimension size of the vector representation. The authors calculated the positive pointwise mutual information (PPMI) in each cell, considering the words and their co-occurrences. Although the vocabulary (rows) can be updated, similarly to bag-of-words, if the contexts are fixed, the system may incur obsolescence after the context words decrease or stop appearing. Furthermore, if certain context words are exchanged with other words, the changed dimensions will not represent the same contexts, and this will be reflected in an ML model dependent on vector inputs.

In this subsection, we analyzed the text representation methods used in the selected papers. However, we do not extrapolate the same analyses to incremental versions. Thus, when we discuss that a particular method only works at least in batches, we do not extend the same conclusions to other versions, including incremental/adaptive versions when available.

We list text representation methods with incremental update capabilities organized in windows/batches or instance. Considering the update in windows/batches, this indicates that the text representation method requires a new amount of data to either be worth updating or satisfy a specific constraint of the text representation method. Using BERT as in Amba Hombaiah et al. (2021) and Susi and Shanthi (2023) are examples of this category. In Amba Hombaiah et al. (2021), the BERT model is fine-tuned using texts selected by the sample methods proposed by the authors. In Susi and Shanthi (2023), the fine-tuning is performed through an updated training set. The training set is updated whenever a sentiment drift is deemed to have occurred.

Considering the incremental methods that can be updated in instances, it implies that it is unnecessary to accumulate data to update the text representation method: a single piece of information can be used for that. For example, we mention Incremental Word Context (Bravo-Marquez et al., 2022). Furthermore, given a single new input, the Graph-of-Words (Yang et al., 2021) can be updated in real time.

Considering the text representation methods that do not allow any update but are retrained from scratch, we list bag-of-words, bigrams, biterm, and TF-IDF. However, while in use, a few text representation methods are kept static in the text streams: FastText, Doc2Vec, and GloVe. Most are used as pre-trained models, and they can become obsolete after some time, demanding complete retraining to maintain the performance of the dependent ML model. Works such as Heusinger et al. (2022); Vo (2022) and Li et al. (2022) also leveraged static BERT and Word2Vec models.

This dataset contains approximately 20,000 news across 20 groups. In the link provided in this paper, there are three versions of this dataset, with slight variations.

According to Lu et al. (2022), this dataset contains approximately 2 million abstracts of papers published comprising the years between 2007 and 2021.

Pohl et al. (2018) cited that CrisisLexT26 (Olteanu et al., 2015) is a collection of datasets related to several crises worldwide. However, Pohl et al. (2018) used only the datasets related to the Colorado Floods, containing 751 relevant and 224 irrelevant tweets, and Australian Bushfires, containing 645 relevant and 408 irrelevant tweets.

This dataset contains 1500 samples with 913 dimensions, i.e., boolean bag-of-words, corresponding to email messages, to be classified as junk or interesting. According to Katakis et al. (2010), these samples were collected from Usenet posts existing inside the 20Newsgroup dataset.

EveTAR is an Arabic dataset that contains 1392 tweets on three terrorist events: (i) a suicide bombing in Ab, Yemen; (ii) Air strikes in Pakistan; and (iii) the Charlie Hebdo attack. This dataset is used in Abid et al. (2019) to evaluate the ability of AIS-Clus to receive texts and detect events in languages other than English.

Wang et al. (2019) collected a news stream from The Guardian using the API. The dataset contains 10 categories and 40,000 samples, represented using Word2Vec with 300 dimensions.

The Irish Times dataset corresponds to a set of 1.6 million news headlines published by the Irish Times, distributed in six classes. It comprises 25 years of publications.

NELA-GT (Nørregaard et al., 2019; Gruppi et al., 2020, 2021) corresponds to a series of datasets regarding news and media outlets. In addition, conspiracy sources are included in this dataset. The authors incorporated ground-truth ratings of aspects such as reliability, transparency, and bias. NELA-GT-2018 (Nørregaard et al., 2019) contains 713 thousand items from 194 media outlets and conspiracy sites; NELA-GT-2019 (Gruppi et al., 2020) contains 1.12 million media articles from 260 mainstream and alternative sources collected in 2019; NELA-GT-2020 (Gruppi et al., 2021) contains almost 1.8 million news stories from 519 sources.

Fenza et al. (2023) used these datasets in the fake news detection, using the instances labeled as reliable and unreliable. These datasets were merged, but the temporal order was respected.

Wang et al. (2019) used The New York Times’ public API to collect news articles, between January 2006 and January 2018. These news articles were encoded using Word2Vec, with 300 dimensions. He et al. (2018) used a dataset collected from The New York Times, containing news articles from 1987 and 2007, distributed in 26 categories (Sandhaus, 2008). Van Linh et al. (2022) used only the title of news articles from the New York Times. The authors mentioned that the dataset contained 1,764,127 titles, with an average of five words per title. Lu et al. (2022) utilized a dataset collected from the News York Times containing 99,872 articles, dating from 1990 to 2016.

National Oceanic and Atmospheric Administration (NOAA) is an agency in the United States government. It does not correspond directly to a dataset; however, Suprem and Pu (2019a), Suprem et al. (2019b), and Suprem and Pu (2019b) used NOAA reports as ground truth for the automatic classification of tweets. No details were offered about the reports’ processing or collection.

The NSDQ dataset (named after NASDAQ) corresponds to tweets regarding 15 companies listed in NASDAQ. NSDQ was compiled by the authors in Heusinger et al. (2020a) and Heusinger et al. (2022), and comprised the months of February to December 2019. This dataset contains 30,278 tweets.

Amba Hombaiah et al. (2021) used the OffensEval 2019 dataset (Zampieri et al., 2019). The dataset contains 14,000 tweets posted in 2019 divided into offensive and inoffensive.

RCV1 (Lewis et al., 2004) is a dataset that contains 403,143 news from Reuters News between 1996 and 1997. The news articles are divided into three classes: industries, topics, and regions. This dataset is organized hierarchically. From this dataset, He et al. (2018) obtained a corpus with 12 subtrees (labels).

Murena et al. (2018) used this dataset, which contains articles with their respective categories, temporally ordered. According to Murena et al. (2018), it contains 12,902 news, each classified into several categories, totaling 90 categories.

Periti et al. (2022) used the datasets corresponding to Task 2 of SemEval2020, regarding the semantic shift detection task. The datasets used are CCOHA (Alatrash et al., 2020) and LatinISE (McGillivray and Kilgarriff, 2013). CCOHA contains texts in English approximately ranging from 1810 to 2000, while LatinISE has Latin texts, ranging from the 2nd century BC to the 21st century AD. Both have target words, which are words that can be monitored to detect the semantic shift. These datasets are discovered in the selected papers that span the longest.

Rakib et al. (2021) collected duplicated question titles regarding Python, Java, jQuery, R, and other programming languages/tools. In the paper, the authors carefully described the process of obtaining this dataset. In the end, this dataset contained 400,000 randomly selected pairs of question titles.

These datasets correspond to emails collected from the Spam Assassin collection. They are represented as bag-of-words, distributed across two classes: ham and spam, in imbalanced proportions (80% and 20%, respectively). Both contain 9,324 instances; however, SpamAssassin Katakis et al. (2008) has 40,000 features, while SpamData has 499 (Katakis et al., 2009). It is noted that these datasets contain gradual drifts (de Moraes and Gradvohl, 2021).

Yin et al. (2018) used this dataset named Tweets, containing 30,332 tweets distributed into 269 groups, with 7.97 words per tweet on average. The authors also generated a variant dataset from Tweets, called Tweets-T, where the dataset is sorted by topic. Rakib et al. (2021) used the same dataset Tweets, called Ts-T. In Vo (2022), the same datasets presented in Yin et al. (2018) are named Tw and Tw-T, respectively.

Li et al. (2018) used a Tweets dataset containing approximately 400,000 tweets. They stated that the data acquisition comprises November and December 2012, using the Twitter API. The dataset in Sun et al. (2021) is also obtained through Twitter API and consists of 803,613 short texts distributed in four categories. Kolajo et al. (2022) described the dataset used in their work as “Twitter sentiment analysis training corpus”, from which they filtered 10% of the data, totaling 104,857 tweets. D’Andrea et al. (2019) collected tweets by using a Java library named GetOldTweets¹⁴¹⁴14https://github.com/Jefferson-Henrique/GetOldTweets-java/. They collected 112,397 tweets posted between September 2016 and January 2017, using vaccine-related keywords. Bechini et al. (2021) extended the dataset obtained in D’Andrea et al. (2019) until September 2019, corresponding to 806,672 tweets. Bondielli et al. (2022) collected 486,688 tweets from July 2021 to December 2021 regarding the Green Pass, as the European Union Covid-19 Digital Certificate is known in Italy. Amba Hombaiah et al. (2021) used tweets to perform country hashtag prediction, in two different years: 2014 and 2017, consisting of 472,000 and 407,000 tweets respectively. The tweets were obtained from the Internet Archive¹⁵¹⁵15https://archive.org/details/twitterstream. Yang et al. (2021) experimented with their approach using a Twitter dataset, namely TweetSet, containing about 144,000 tweets posted in June 2019, distributed into 16 categories. de Mello et al. (2018) collected tweets by monitoring a set of users and hashtags, i.e., words with a # at the beginning that simulate a tag for the tweet. The authors monitored, for instance, @dilmabr (former Brazilian president) and #dolar (Portuguese for dollar). The dataset size was not mentioned.

Although Twitter-based datasets were very frequent in the studied papers, as of February 2023, Twitter’s API policies have changed¹⁶¹⁶16Available at: https://www.forbes.com/sites/jenaebarnes/2023/02/03/twitter-ends-its-free-api-heres-who-will-be-affected/?sh=36ad308a6266. Accessed on September 17th, 2023., and it became a paid service.

TwitterSentiment (or TSentiment, as in Melidis et al. (2018)) is a balanced dataset that contains 1.6 million tweets collected between April and June 2009. These tweets are labeled as positive or negative using distant supervision. In this case, emoticons were used for labeling.

The UCINews dataset contains 422,937 news collected between March and August 2014. Each news can be categorized as business, science and technology, entertainment, and health. This data collection also includes each news id, title, URL, publisher, story id, hostname, and timestamp information.

Similarly to EmailingList, both Usenet1 and Usenet2 simulate a sequence of 1500 emails from 20NewsGroup dataset to a particular user to be classified as junk or interesting (Katakis et al., 2008). Both datasets have 100 features, corresponding to words, according to de Moraes and Gradvohl (2021).

United States Geological Survey (USGS) is a scientific agency from the United States. Similarly to NOAA, USGS reports do not correspond to datasets and are also used as ground truth to classify tweets automatically in Suprem and Pu (2019a), Suprem et al. (2019b), and Suprem and Pu (2019b).

Vg.no is a Norwegian news website. Hammer and Yazidi (2018) obtained news from four topics: European Union, economy, sports, and entertainment. However, the authors did not mention the size of the collected dataset.

Mohawesh et al. (2021) used four real-world datasets, based on the datasets provided by Yelp, namely Yelp CHI, Yelp NYC, Yelp ZIP, and Yelp Consumer Electronics. The authors used Yelp CHI (Chicago) (Mukherjee et al., 2013), containing more than 67,000 reviews of restaurants and hotels, distributed between 2004 and 2012. Yelp NYC (Rayana and Akoglu, 2015) contains approximately 322,000 reviews of restaurants located in New York City. It comprises the years between 2004 and 2015. Yelp ZIP (Rayana and Akoglu, 2015) contains 608,598 reviews from New Jersey, Vermont, Connecticut, and Pennsylvania. Yelp Consumer Electronics (Barbado et al., 2019) contains almost 19,000 records evenly distributed between genuine and fake. These datasets include other data, such as user information, product information, rating, timestamp, and review, and were scraped/downloaded from Yelp.com.

Although SpamAssassin and EmailingList have known concept drifts (gradual and abrupt)¹⁷¹⁷17According to http://mlkd.csd.auth.gr/concept_drift.html, an interesting aspect is that none of the datasets has labeled concept drifts, due to the difficulty of defining the specific points of drift, which requires a deep study on a particular dataset. Thus, some authors only warned about the unawareness of concept drift in the datasets. However, others attempted to force concept drifts by: (i) placing data partitions temporally disordered in a stream, i.e., data from 2011 and 2015 before 2012 (Mohawesh et al., 2021); or (ii) rearranging the data, sorting by classes or topics (Li et al., 2018). This aspect is extended in Section 6.

Therefore, since we could not locate repeating datasets in more than three papers, we can conclude that the research area of concept drift detection in textual streams lacks benchmark datasets. Furthermore, all the datasets used for classification are instance-level labeled, i.e., sentences/tweets labeled. In addition, the resource of one of the most recurrent datasets, i.e., TagMyNews and Snippets (Phan et al., 2010), could not be encountered across the papers. Also, it is closely related to short-text applications, which constitutes an entirely new research area.