© 2024 JETIR November 2024, Volume 11, Issue 11 www.jetir.

org (ISSN-2349-5162)

LEARNING UNDER CONCEPT DRIFT: A

LITERATURE REVIEW
1 Ashwini M C, 2 Dr. Minavathi,3 Hardik Jain
1 Asst.Professor 2 Professor and Head, 3 Research student
1Department of Computer Science and Engineering,
1PESCE, Mandya, India
Abstract: In recent years, the availability of time series streaming information has been growing enormously. Learning from real-
time data has been receiving increasingly more attention since the last decade. Online learning encounters the change in the
distribution of data while extracting considerable information from data streams. Hidden data contexts, which are not known to
the learning algorithms, are known as concept drift. Classifier classifies incoming instances using past training instances of the
data stream. The accuracy of the classifier deteriorates because of the concept drift. The traditional classifiers are not expected to
learn the patterns in a non-stationary distribution of data. For any real-time use, the classifier needs to detect the concept drift and
adapts over time. In the real-time scenario, we have to deal with semi-supervised and unsupervised data, which provide no or
fewer labeled data. The motivation behind this paper is to introduce a survey identified with a broad categorization of concept drift
detectors with their key points, limitations, and advantages. Eventually, the article suggests research trends, research challenges,
and future work. The adaptive mechanisms are also incorporated in this survey.

Index Terms - drift data streams, natural language processing, sentiment analysis, sentiment drift.

LEARNING UNDER CONCEPT DRIFT: ALITERATURE REVIEW

I. INTRODUCTION
Sentiment analysis, involves gathering people’s opinions, such as their positive and negative views and emotions, through
natural language processing and text mining [1]. As a technique, NLP numerically gives emotional polarity scores of the concepts
found in the texts where they have been utilized, such as newspapers and social media. On text data from online, models use machine
learning vocabulary dictionaries, graph- based learning, and others to classify and predict emotional values [3]. Sentiment drift
occurs when the distribution of sentiment is underlined in a way most of the previously learned patterns are known to be ineffective
[4]. Contextual mining facilitates the examination of articulated concepts, taking into account not solely the textual indicators but
also the structural characteristics of language and linguistic patterns to effectively capture sentiment. [5]. Nevertheless, people
created difficulties for machines as their prompts are more complicated with many aspects such as sarcasm, irony, etc., which creates
many challenges, allowing a machine to misinterpret [6]. If there are biases in collecting and utilizing opinion datasets and collecting
and annotating data, then even machine learning algorithms might be implicated in bias [7]. Bias discovers a way by observing the
data and models concerning the demographic, social, or ideological angle. Users are clustered into focused and diverse groups to
reduce the bias in user peer acceptance caused by users’ different interest diversity [8]. This implies that investigations of genetic
material may require getting consent from the users, deciding the right policy for collecting and safeguarding data, and revealing
the results and potential by-products. Ethical
frameworks applicable in opinion mining are often developed to achieve universal acceptance or cross-authority compliance
among all parties involved. Also, as technologies and social norms tend to change over time, these frameworks are reviewed over
time for changes [10].

JETIR2411319 Journal of Emerging Technologies and Innovative Research (JETIR) www.jetir.org d163
© 2024 JETIR November 2024, Volume 11, Issue 11 www.jetir.org (ISSN-2349-5162)

Fig. 1. General block diagram of concept drift detection.

The remaining section of this paper discusses related works on opinion mining in drift data streams in section 2. A taxonomy of
Opinion mining in Drift data streams is presented in section 3. A comparative analysis of various approaches is discussed in section
4. Section 5 discusses the results of different algorithms and section 5 represent a summary of this paper.
II. LITERATURE SURVEY
Sanjeev Kumar et al. [11] introduced a novel Adaptive Ensemble Classifier (AEC) for online sentiment analysis and opinion
mining. This unique approach utilizes false-positive drift detections to mitigate the negative effect of valuable signals and false-
positive drift detection. Whenever a drift occurs, a new classifier is seamlessly integrated into the ensemble. A weighting technique
assigns weights to each classifier, regulating their involvement in the final classification outcomes. Despite the longer than projected
analysis time on high-dimensional text datasets, the AEC shows promise in its innovative approach. This paper presents the
classification accuracy results of the various models tested on four different datasets. The results indicate that the Drift Detection-
based Adaptive Ensemble (DAE) classifier consistently outperformed single classifiers across all datasets. Specifically, the DAE
model achieved the highest accuracy on the Amazon dataset, surpassing 76.5%, which was attributed to its effective handling of
concept drift. This paper also highlights that the best classification accuracy was obtained when the warning significance level was
set to 0.3 and the detect significance level to 0.09, resulting in five detected drifts. This configuration allowed the DAE model to
maintain high accuracy while managing the number of false-positive drift detections. In comparison, traditional models like OZA,
which employs ADWIN for drift detection, did not achieve similar accuracy levels. The results underscore the DAE model's
robustness in adapting to changes in data streams, making it a valuable approach for sentiment analysis and opinion mining tasks.
Li Yang and Abdallah Shami [12] developed the detection and adaptation structure of lightweight concept drift. Adaptive IoT
streaming data analytics was developed based on optimized LightGBM and concept drift adoption for anomaly detection. The new
adaptation approach in drift, Optimized Adaptive and Sliding Windowing (OASW), was developed to adapt to pattern variations of
online IoT data streams. The high speed in data processing exhibits the probability of implementing the developed architecture in
real-time environments on powerful cloud machines. OASW detected the concept drift and updated the learning model mainly based
on the model's performance degradation, which assures the learner only updates when essential. The proposed method can achieve
the highest accuracy of 99.92% among all implemented models by adapting a slight concept drift detected at point 13408. Without
drift adaptation, the offline LightGBM model has a slightly lower accuracy of 99.78%. The accuracies of the other four state-of-
the-art methods are also lower than the accuracy of our proposed method (99.01%- 99.27%).
Hyejin Jang et al. [13] suggested exploring approach opportunities based on user requirements for applying opinion mining and
subject action object (SAO) analysis. This approach was developed to explore technology opportunities that analyze the SAO
frameworks discovered in patents and user reviews. Then, patent documents were arranged regarding technological elements
presented in patent analysis. The SAO frameworks of technologies were extracted by utilizing their F-term code, which comprises
the multidimensional data in this classification system. SAO frameworks were vectorized utilizing SAO2Vec before computing
cosine similarity among SAOs to link reviews and patents obtained from reviews and patent documents. However, the SAO
approach, developed by verifying the direction of technology development in this research, must verify actual industry application.

JETIR2411319 Journal of Emerging Technologies and Innovative Research (JETIR) www.jetir.org d164
© 2024 JETIR November 2024, Volume 11, Issue 11 www.jetir.org (ISSN-2349-5162)
The study identified a total of 9,954 patents related to "electric vehicles," which was the highest number found in a specific area of
automotive-related needs. This indicates a significant focus on electric vehicle technology within the patent landscape
Abdulaziz O. AlQabbany and Aqil M. Azmi [14] introduced the efficiency of Adaptive Random Forest (ARF) in data streams
to manage concept drift. This study aims to investigate the problem of concept drift in random forests and, hence, come up with an
approach that can be utilized to develop an ideal prediction model that combines knowledge acquisition techniques. In such cases,
where circular reasoning occurs within a large data set, it is possible to approximate the technique utilizing Poisson distribution.
The ARF algorithm exhibits capable outcomes for handling various types of drift and is accurate. Specific approaches were applied
to develop an implementation for the random forest algorithm to demonstrate its performance in processing three different data sets
in two languages. However, a developed approach was required to analyze the concept drift phenomenon with a sophisticated Big
data analytics framework. The best average accuracy achieved across the datasets was 89.88%, which occurred at λ values of 8 and
10. However, this high accuracy was associated with a significant execution time of 346 seconds for λ = 8. In a specific case study,
both λ = 3 and λ = 6 yielded an accuracy of 79.7%, but the execution times differed: 17 minutes and 45 seconds for λ = 3 and 28
minutes and 8 seconds for λ = 6. This indicates a speedup of approximately 1.585 times when using λ = 3 over λ = 6.
Anna Koufakou [15] suggested deep learning for opinion mining and topic classification of course reviews. The numerous
course reviews were collected pre-processed and publicly available online. The machine learning approaches were applied to gain
insight into student sentiments and topics. The utilization of current NLP approaches such as Generalized Auto-regression Pre-
training (XLNet), word embeddings and deep neural networks, robustly optimized BERT approach (RoBERTa), state-of-the-art
Bidirectional Encoder Representations from Transformers (BERT). Extensive experimentation was performed to compare these
techniques to traditional techniques. However, the advanced approach was required to analyze aspect-based sentiment analysis. The
top-performing model for sentiment polarity extraction was RoBERTa, achieving an accuracy of 95.5% and an F1-macro score of
84.7%. This indicates that RoBERTa was highly effective in determining the sentiment expressed in course reviews. For the topic
classification task, the best-performing model was the Support Vector Machine (SVM), which reached an accuracy of 79.8% and
an F1-macro score of 80.6%. This suggests that while deep learning models were explored, traditional models like SVM still
performed better in this specific task.
III. TAXONOMY
T In this section, we present a taxonomy for opinion mining in drift data streams utilizing various approaches such as Adaptive
Learning, Ensemble Learning, Machine Learning, and Deep Learning. These categories comprise the various techniques for opinion
mining in user-based drift data streams and provide better outcomes.
Figure 1 represents the taxonomy diagram of opinion mining in drift data streams.

Fig. 1. Taxonomy Diagram of opinion mining in drift data streams

A. Adaptive Learning
Sentiment analysis through adaptive learning has shown excellent performance in natural language processing. Different
Adaptive learning techniques are utilized in Twi, such as Sentiment Drift Analysis–Bidirectional Encoder Representation and
Transformer. This algorithm can be described in the following sub-section.
1) Twitter Sentiment Drift Analysis–Bidirectional Encoder Representation and Transformer: E. Susi and A. P. Shanthi [16] developed
the detection and analysis of sentiment drift in real-time Twitter data streams. The dynamic drift measure is a TSDA- BERT feature
introduced earlier. In contrast, the second one included the domain impact score as an update component that couldleave a mark on
the input domain-specific data. It has an adaptive learning strategy for identifying sentiment shifts in microblog

JETIR2411319 Journal of Emerging Technologies and Innovative Research (JETIR) www.jetir.org d165
© 2024 JETIR November 2024, Volume 11, Issue 11 www.jetir.org (ISSN-2349-5162)

posts through a multithreading attention mechanism and utilizing the BERT structure. The solution can be dynamic too for a chore
of Kafka tool to proceed with this transformation by converting static data streams into moving streams platforms. It accurately
recognizes sentiment variation swings, finds and processes their results automatically, and detects sentiment in real-time data
streams.
2) AiRStream: Ben Halstead et al. [24] suggested the analyzing as well as repairing concept drift adaptation in classification of data
stream. This research was developed based on a new algorithm involving both the detection and fixing errors in the adjustments
made due to the concept drift. This evaluation technic based on artificial data had exhibit that AiRStream approach gives better
results compared to baseline methods, being more apropos for catching of dynamical streams. In addition, the inference task of
the clean air was successful in reality, where AIRStream shows better performance than the eight baseline methods. Modeling the
concept of drifting with AiRStream promises advanced inferences of the environment parameters in uncharted regions.

B. Ensemble Learning
Ensemble learning is a powerful machine learning technique that combines the predictions from multiple individual models to
achieve better predictive performance than any single model. Different types of Ensemble learning techniques are utilized, suchas
Streaming Classification based on Hierarchical Concept Drift and Online Ensemble, transformer, and ELStream. This algorithmcan
be described in the following sub-section.
1) Streaming Classification based on Hierarchical Concept Drift and Online Ensemble: Ning Liu and Jianhua Zhao [17] presented
the Classification of streaming data based on the hierarchical concepts of drift and online ensemble. The concept drift index is
produced utilizing incoming data instances. It divides the streaming data into three states: well-witness concept drift alerting and
the appearance of a concept drift phenomenon. During this time, the classifier learns nothing new because no additional examples
have yet to be given. During the action drift warning phase, online ensemble learning is utilized for feature extraction and model
upgrading with superior efficiency. Here, while an anomaly detection mechanism catches anomalous data and strikes them out,
hybrid ensemble online learning combined with incremental learning is utilized for adaptation. The model directs the robot to wards local
knowledge sources from the data stream when the occurrence of idea drift events is detected.
2) ELStream: Abbasi et al. [18] suggested the ensemble learning technique in dynamic social significant data stream learning for
concept drift detection. The research proposes ElStream, an exciting approach involving ensemble and traditional machine learning
models to identify both natural and artificial concept drifts. The same voting approach that ElStream applies, which implemented
the abstaining technique, as only the first classifier in the whole competition is allowed to make the final decision. Performance
shows that the ElStream outperforms the traditional machine learning ones by accuracy and f-scores, whereas the Ensemble
classifier has the best overall accuracy.

C. Deep Learning
Deep learning performs a vital part in sentiment drift data analysis, allowing to uncover patterns and trends within textual
data. Different types of deep learning techniques are utilized for Hybrid Uncertainty Calibration and Class Imbalance with Drift
Detection – Adadelta Optimizer based Deep Neural Network. This algorithm can be described in the following sub-sections.
1) Hybrid Uncertainty Calibration Approach: Qiuyu Pan and Zuqiang Meng [19] presented the hybrid uncertainty calibration for
multimodal sentiment analysis. To begin with, this research proposes to investigate the task of sentiment polarity distribution in
MSA data sets with the ultimate aim of providing an uncertain-aware fusion approach that would give a smooth fusion modality
of modalities. Second, this paper discusses a hybrid uncertainty calibration algorithm built within evidential deep learning (EDL)
context to provide better calibration, give more certainty, and reduce the uncertainty in each model's modalities. Lastly, it will
apply the simulated noise and verify the performance of the design noise. Unlike the benchmark approaches, this approach
performs best in classification and calibration performances on three MVSA datasets.
2) Class Imbalance with Drift Detection – Adadelta Optimizer based Deep Neural Network: S. Priya and R. Annie Uthra [23]
presented the DL approach to handle concept drift as well as class imbalanced complex decision making on streaming data. For
classification in highly imbalanced streaming data, an efficient class imbalance over detection of concept drift was presented by
Adadelta optimizer-based deep neural networks (CIDD-ADODNN). The developed approach utilizes adaptive synthetic approach
for dealing class imbalance data that applies weighted distribution for various minority class. Then, the adaptive sliding window
from drift detection approach was used for detecting the concept drift presence. An appropriate ADO based parameter tuning
approach was used to locate the best DNN technique parameters. Outcomes of simulation agreed with the prior stated conclusion
of the model by achieving a maximal accuracy degree.

D. Machine Learning
Machine learning also plays a crucial role in sentiment drift data analysis, allowing us to uncover patterns and trends within
textual data. Machine learning techniques can be applied to handle concept drift in opinion mining. Different types of Machine
learning techniques are utilized, such as Peer Acceptance, Co-Training, Self-Learning, Expectation Maximization, and
Adversarial Dual Discriminator. These algorithms can be described in the following sub-section.
1) Peer Acceptance: Koggalahewa et al. [20] introduced the unsupervised approach for detection of social network spammer based
on user data interests. This analysis presented a fully unsupervised technique of spammer detection, relying on mutual agreement
among members in a social network as a means to differentiate between regular users and spammers. The concept of peer
acceptance is formed through analyzing user-specific common interests in many topics. The main innovation described in this
paper is the idea of an unsupervised technique for spammer detection utilizing users’ peer acceptance that does not require the
labeled data. Also, since it is unsupervised, this approach does not need a new training model to be built if data changes.
2) Co-Training, Self-Learningand Expectation- Maximization: Io Vasileios Iosifidis and Eirini Ntoutsi [21] implemented the

JETIR2411319 Journal of Emerging Technologies and Innovative Research (JETIR) www.jetir.org d166
© 2024 JETIR November 2024, Volume 11, Issue 11 www.jetir.org (ISSN-2349-5162)

sentiment analysis on significant, sparse data streams with limited labels. This research aims to analyze the insights gained through
our annotation process and the results related to 3 distinct semi-supervised learning approaches: self-learning, peer training, and
expectation maximization. In addition, this research suggests ways of annotation: whether algorithms keep all classified and
unlabeled statistics in mind or follow the sequence of circulating statistics by processing them through a lightweight streaming
mode. Also, inside the magnitude disparity problem, this phenomenon is exacerbated when utilizing semi-supervised strategies,
as this study finds that in our asymmetric dataset, the most significant sense class is the majority.
Adversarial Dual Discriminator: Siddique Latif et al.[22] presented the self-supervised adversarial domain adaptation for cross-
corpus and cross-language speech emotion recognition. This research introduces the Adversarial dual discriminator (ADDi)
community, which utilizes a 3-player hostile game to examine generalized representations without requiring target statistics labels.
It additionally includes a self-supervised variant of the ADDi network (sADDi), which employs self-supervised pre- schooling on
unlabeled input. This look proposes utilizing synthetic facts epoch as a pretext mission in sADDi, allowing the network to as semble
emotionally discriminative and area- invariant representations and providing extra artificial facts to improve the system. A
developed version has severely assessed the usage of publicly accessible datasets in three languages and contrasted it with other
Speech Emotion Recognition (SER) research.

IV. COMPARATIVE ANALYSIS

The opinion mining in drift data streams is compared with existing models to enhance their performance. The comparative
analysis is significant in developing and effectively enhancing the model's performance. The comparative analysis comprises the
employed methodology, advantages, limitations, and performance measures. Table 1 represents the comparative analysis of existing
approaches
TABLE I. COMPARATIVE ANALYSIS OF EXISTING APPROACHES
Sl No Author Methodology Employed Advantages Limitations Observations
1 Susi and The detection and analysis Adaptive learning of However, this approach focused on The developed model uses
Shanthi [16] of sentiment drift is approach accomplished exploring the possibility of adaptive learning to be more
developed in in twitter significantly better extensions for filtering tweets from accurate even when sentiment
data streams by utilizing accuracy through sentiment bot as well as fake accounts and drift change. It
TSDA-BERT drifts. enhance the performance of the also reduce consumption of time.
structure.
2 Liu and Zhao The classification of A learning modelresponded Additional optimization of algorithm This approach enhanced the
[17] streaming data is quickly afterconcept was required to consider expanding learning efficacy though maintain
presented based on While enhancing its applicability though additional the learning accuracy.
hierarchical concept drift robustness. eliminating time costs.
and online ensemble
3 Abbasi et al. The online ensemble This result suggestively The developed approach was The confidence level can only be
[18] learning approach is aided for the application of required to diminish the predicted by each classifier in the
suggested in dynamic big data analytics in real- computational complexity of this group. The Elstream approach
social big data stream life applications. algorithm. provided higher accuracy and F-
learning for identifying score ratios compared to similar
concept drifts state-
of-the-art studies.

4 Pan and An uncertainty-aware Using ULF-HUC together The developed approach was When the model is accurate, a
Meng [19] delay fusion based on made the approach better. It required to consider more appropriate learning method effectively
hybrid uncertainty also made model fusion approaches for calibrating learning reduces uncertainty with
calibration is developed more reliable and reduced abilities to solve the difficulties of hybrid uncertainty
for multi-model sensory uncertainty. disparate learning capabilities calibration.
analysis. between various modalities.
5 Koggalahewa An unsupervised The approach did not need The probability of such a situation Detection of spam based on peer
et al. [20] approach is developed a labeled data set. Because was very less when the number of acceptance can augment
based on user data it was unsupervised, a topics utilized was high detection
interests for social novel training model was accuracy by integration of
network spammer not essential when data clustering
detection changed.
6 Iosifidis and Sentiment analysis is Semi-supervised learning However, this approach was required AU-PRC better reflects the
Ntoutsi [21] presented on big sparse and data augmentation to analyze the additional data performance and accuracy of the
data streams with limited suggestively outperformed augmentation approaches as a tool to classifier, and offers more useful
labels for addressing the the default semi-supervised expand the training set for tackling representations than
class imbalance issues annotation procedure. the class imbalance problems AUC.

7 Latif et al. The adaptation of self- A self-monitoring pretext However, SER was not a stable When single or multiple
[22] supervised adversarial task produced artificial data representation for the transfer task, in components are
domain is developed for as a byproduct to enhance spite of auto-encoder popularity eliminated, this approach has a
cross-language and cross- system for accomplishing based reconstruction excuse task in significant drop in UAR
corpus speech emotion ideal performance. computer vision.
recognition
8 S. Priya and A DL approach is The simulation results The performance of CIDDO- The novel ADODNN, also
R. Annie developed to handle confirmed the presented ADODNN model needed to be CIDD-ADODNN
Uthra [23] complex decision-making model's exceptional enhanced by feature selection and approaches have obtained
with concept drift and class performance, achieving clustering approaches to solve the ideal functions at gaining accurate
imbalance in streaming data. maximum accuracy. problems of accurate drift detection, outcomes.
class disparity and efficient
hyperparameter tuning process.

JETIR2411319 Journal of Emerging Technologies and Innovative Research (JETIR) www.jetir.org d167
© 2024 JETIR November 2024, Volume 11, Issue 11 www.jetir.org (ISSN-2349-5162)

9 Ben Halstead The analyzing as well as The AiRStream concept drift This approach is required to utilize The idea drift model produced by
et al. [24] restoring adaptation of concept model holds promise for variations spotted by AiRStream for AiRStream can be used to predict
drift in classification from data inferring environmental enhancing interpretation performance wind speed and direction
stream conditions in previously and addressing the affected by with high accuracy.
inaccessible locations. variation in unknown features

V. EXPERIMENTAL RESULTS AND DISCUSSION

This section presents the DDAW algorithm by employing four different distance measures and then compare the results with
the baseline approaches using the two real-world datasets namely, Amazon shopping and 20-Newsgroup [25]

5.1 Investigation into the Performance of Concept Drift Detection Methods

Table 1 summarizes the experimental findings for three baseline methods and our proposed method that employs four distinct
dissimilarity measures, for a total of ten (10) different approaches in terms of false alarm (FA) and missing rate (MR). Then, Figs.
3 and 4 exhibit the performance of drift detection methods on the Amazon shopping and 20-News group datasets for more clarity.
The DDAW-HD model is superior to the other DDAW versions and the three based lines on both datasets, as shown in Table 1

Fig 2: Comparison of drift detection methods with DDAW versions based on the false alarm on Amazon and 20-Newsgroup
datasets

Fig 4: Comparison of drift detection methods with DDAW versions based on miss rate on Amazonand 20-Newsgroup
datasets

5.2 Investigation into the Classification Results with Various Drift Detection Methods

JETIR2411319 Journal of Emerging Technologies and Innovative Research (JETIR) www.jetir.org d168
© 2024 JETIR November 2024, Volume 11, Issue 11 www.jetir.org (ISSN-2349-5162)

As demonstrated in Table 2, the average error rate relates to the average of all data chunks as well as the standard deviation for
those error rates. In terms of average accuracy, our DDAW-HD method surpasses previous methods on both the 20-Newsgroups
and Amazon shopping datasets.

Table 2: Effect of classifiers on the drift detectio

Classifier DDM EDDM Page-
hink

NB 0.25 ± 0.25
SVM 0.2

Based on Table 2, the DDAW-HD approach minimizes the error rate in 20-Newsgroups to 26.76%, which significantly
outperforms the alternatives. This is due to the fact that DDAW-HD has fewer missed drifts and false alarms, resulting in a lower
classification error. When DDAW-HD detects a concept drift, rather than recreating the model on the most recent chunk, it will
incrementally train the classifier and make a prediction based on its incoming data, resulting in a very low classification error. On
Amazon, where concept drift is uncommon, DDAW-HD performs similarly to baselines. The DDAW-HD method outperforms
the competing algorithms in every scenario.

Fig 5: Concept drift detection curves on the 20-Newsgroups dataset

Fig 6: Concept drift detection curves on Amazon shopping dataset

JETIR2411319 Journal of Emerging Technologies and Innovative Research (JETIR) www.jetir.org d169
© 2024 JETIR November 2024, Volume 11, Issue 11 www.jetir.org (ISSN-2349-5162)

VI. CONCLUSION
In conclusion, for data streaming and sentiment analysis, remodeling for drift is critical because it immediately affects the
accuracy and value of sentiment analysis models. It is significant to detect and respond to sentiment drift in statistical streams to
maintain the general performance of the sentiment analysis model. The existing opinion-mining approaches utilize temporal
sentiment analysis to analyze how sentiment intensity changes and drifts emerge across various periods and aid businesses in
understanding and accomplishing their product status. Further, numerous analysis frameworks are utilized to estimate the
performance of concept mining models on dynamic data streams, which considers metrics that enumerate temporal aspects, such as
accuracy over time, the correctness of concept detection, or the modality of concepts. However, assessing sentiment analysis in
multilingual contexts poses hurdles, as translating and labeling sentiment across multiple languages can add errors and
inconsistencies due to place-unique tones and linguistic variations.
The paper suggests that future work could focus on enhancing the drift detection mechanisms to make them more sensitive and
accurate. This could involve developing more sophisticated algorithms that can better identify subtle changes in data streams,
thereby improving the overall performance of classifiers in dynamic environments. As the paper discusses the potential for real-
time applications, future work could focus on optimizing the algorithms for faster processing speeds. This would make the models
more applicable in real-time environments, particularly in Twitter data streaming data scenarios. the future enhancements of the
paper could focus on improving drift detection mechanisms, integrating additional data sources, optimizing for real-time processing,
exploring hybrid models, addressing ethical considerations, and conducting longitudinal studies to ensure the robustness and
applicability of the proposed methods.

REFERENCES
[1] Zhang, Y., 2021. Analysis and Recognition Method of Internet Image Public Opinion Based on Partial Differential Equation.
Advances in Mathematical Physics, 2021, p.9759199, 2021. doi: https://doi.org/10.1155/2021/9759199
[2] Bechini, A., Bondielli, A., Ducange, P., Marcelloni, F. and Renda, A., 2021. Addressing event-driven concept drift in twitter
stream: a stance detection application. IEEE Access, 9, pp.77758-77770. doi: 10.1109/ACCESS.2021.3083578
[3] Heusinger, M., Raab, C. and Schleif, F.M., 2022. Dimensionality reduction in the context of dynamic social mediadata
streams. Evolving Systems, 13(3), pp.387-401. https://doi.org/10.1007/s12530-021-09396-z
[4] Abiola, O., Abayomi-Alli, A., Tale, O.A., Misra, S. and Abayomi-Alli, O., 2023. Sentiment analysis of COVID-19 tweets
from selected hashtags in Nigeria utilizing VADER and Text Blob analyser. Journal of Electrical Systems and Information
Technology, 10, p.5. https://doi.org/10.1186/s43067-023-00070-9
[5] Ligthart, A., Catal, C. and Tekinerdogan, B., 2021. Analyzing the effectiveness of semi-supervised learning approaches for
opinion spam classification. Applied Soft Computing, 101, p.107023. https://doi.org/10.1016/j.asoc.2020.107023
[6] Bravo-Marquez, F., Khanchandani, A. and Pfahringer, B., 2022. Incremental word vectors for time-evolving sentiment
lexicon induction. Cognitive Computation, 14(1), pp.425–441. https://doi.org/10.1007/s12559-021-09831-y
[7] Camacho, K., Portelli, R., Shortridge, A. and Takahashi, B., 2021. Sentiment mapping: point pattern analysis of sentiment
classified Twitter data. Cartography and Geographic Information Science, 48(3), pp.241-257.
https://doi.org/10.1080/15230406.2020.1869999
[8] Gopi, A.P., Jyothi, R.N.S., Narayana, V.L. and Sandeep, K.S., 2023. Classification of tweets data based on polarity utilizing
improved RBF kernel of SVM. International Journal of Information Technology, 15(2), pp.965-980.
https://doi.org/10.1007/s41870-019- 00409-4
[9] Zervoudakis, S., Marakakis, E., Kondylakis, H. and Goumas, S., 2021. OpinionMine: A Bayesian-based framework for
opinion mining utilizing Twitter Data. Machine Learning with Applications, 3, p.100018.
https://doi.org/10.1016/j.mlwa.2020.100018
[10] Tu, H.T., Phan, T.T. and Nguyen, K.P., 2022. Modeling information diffusion in social networks with ordinary linear
differential equations. Information Sciences, 593, pp.614-636. https://doi.org/10.1016/j.ins.2022.01.063
[11] Kumar, S., Singh, R., Khan, M.Z. and Noorwali, A., 2021. Design of adaptive ensemble classifier for online sentiment analysis
and opinion mining. PeerJ Computer Science, 7, p.e660. https://doi.org/10.7717/peerj-
cs.660
[12] Yang, L. and Shami, A., 2021. A lightweight concept drift detection and adaptation framework for IoT data streams. IEEE
Internet of Things Magazine, 4(2), pp.96-101. doi: 10.1109/IOTM.0001.2100012
[13] Jang, H., M.S., Park, S. and Yoon, B., 2023. Exploring technology opportunities based on user needs: application of opinion
mining and SAO analysis. Engineering Management Journal, 35(3), pp.209-222.
https://doi.org/10.1080/10429247.2022.2050130
[14] AlQabbany, A.O. and Azmi, A.M., 2021. Measuring the effectiveness of adaptive random forest for handling concept drift in
big data streams. Entropy, 23(7), p.859. https://doi.org/10.3390/e23070859
[15] Koufakou, A., 2024. Deep learning for opinion mining and topic classification of course reviews. Education and Information
Technologies, 29(3), pp.2973-2997. https://doi.org/10.1007/s10639- 023-11736-2
[16] Susi, E. and Shanthi, A.P., 2023. Sentiment Drift Detection and Analysis in Real Time Twitter Data Streams. Computer
Systems Science & Engineering, 45(3), 3231-3246. DOI: 10.32604/csse.2023.032104
[17] Liu, N. and Zhao, J., 2023. Streaming data classification based on hierarchical concept drift and online ensemble. IEEE
Access, 11, pp.126040-126051. doi: 10.1109/ACCESS.2023.3327637
[18] Abbasi, A., Javed, A.R., Chakraborty, C., Nebhen, J., Zehra, W. and Jalil, Z., 2021. ElStream: An ensemble learning approach
for concept drift detection in dynamic social big data stream learning. IEEE Access, 9, pp.66408-66419. doi:
10.1109/ACCESS.2021.3076264
[19] Pan, Q. and Meng, Z., 2024. Hybrid Uncertainty Calibration for Multimodal Sentiment Analysis. Electronics, 13(3), p.662.

JETIR2411319 Journal of Emerging Technologies and Innovative Research (JETIR) www.jetir.org d170
© 2024 JETIR November 2024, Volume 11, Issue 11 www.jetir.org (ISSN-2349-5162)

https://doi.org/10.3390/electronics13030662
[20] Koggalahewa, D., Xu, Y. and Foo, E., 2022. An unsupervised method for social network spammer detection based on user
information interests. Journal of Big Data, 9, p.7. https://doi.org/10.1186/s40537- 021-00552-5
[21] Iosifidis, V. and Ntoutsi, E., 2020. Sentiment analysis on big sparse data streams with limited labels. Knowledge and
Information Systems, 62(4), pp.1393-1432. https://doi.org/10.1007/s10115-019- 01392-9
[22] Latif, S., Rana, R., Khalifa, S., Jurdak, R. and Schuller, B.W., 2023. Self-supervised adversarial domain adaptation for cross-
corpus and cross-language speech emotion recognition. IEEE Transactions on Affective Computing, 14(3),
1912-1926. doi: 10.1109/TAFFC.2022.3167013
[23] Priya, S. and Uthra, R.A., 2023. Deep learning framework for handling concept drift and class imbalanced complex decision-
making on streaming data. Complex & Intelligent Systems, 9(4), pp.3499-3515.
[24] Halstead, B., Koh, Y.S., Riddle, P., Pears, R., Pechenizkiy, M., Bifet, A., Olivares, G. and Coulson, G., 2022. Analyzing and
repairing concept drift adaptation in data stream classification. Machine Learning, 111(10), pp.3489-3523
[25] Rabiu, I., Salim, N., Nasser, M., Da'u, A., Abdalla, T., Eisa, E., Elnour, M., & Dalam, E. (n.d.). Drift detection method using
distance measures and windowing schemes for sentiment classification. https://doi.org/10.32604/cmc.2023.035221

JETIR2411319 Journal of Emerging Technologies and Innovative Research (JETIR) www.jetir.org d171