1 Introduction
Recommendation and search systems have already become indispensable means of helping web users identify the most relevant information/services in the era of information overload. The applications of such systems are multi-faceted, including targeted advertising, intelligent medical assistance, and e-commerce, and are bringing immense convenience to people’s daily lives. However, despite rapid advances in recommendation and search, the increasing public awareness of the trustworthiness of relevant recommendation and search applications has introduced higher expectations on relevant research. First, the unprecedentedly growing heterogeneity of use cases has been challenging the adaptivity of contemporary algorithms to various settings, e.g., dynamic user interests [Chen et al.
2019], highly sparse interaction records [Chen et al.
2020b], and limited computing resources [Long et al.
2022; Imran et al.
2022]. Second, in a broader sense, a trustworthy recommendation/search approach should also be robust, interpretable, secure, privacy-preserving, and fair across different use cases. Specifically, robustness evaluates a model’s performance consistency under various operating conditions like noisy data [Zhang et al.
2020]; interpretability and fairness respectively evaluate if a model can make its decision processes transparent [Chen et al.
2020c;,
2021; Lyu et al.
2021; Cui et al.
2022; Ren et al.
2021] and the decision outcomes unbiased [Chen et al.
2020a; Li et al.
2021; Yu et al.
2022; Yin et al.
2012]; while security and privacy respectively emphasize a model’s ability to handle cyber-attacks [Zhang et al.
2021b;,
2022b; Yuan et al.
2023a] and to prevent personal information leakage [Zhang and Yin
2022; Zhang et al.
2021c;,
2021a; Yuan et al.
2023c; Wang et al.
2022a; Yuan et al.
2023b]. Consequently, trustworthiness is becoming a key performance indicator for state-of-the-art recommendation and search approaches. In light of these emerging challenges, this special section focuses on novel research in this field with the notion of trustworthiness. The articles presented in this special issue will further promote responsible AI applications, thus better universalizing the advanced techniques to a wider range of the common public.
2 Overview of Articles
The submission deadline of the special issue was 15th June, 2022 and we received 41 valid submissions in total; among them, 23 submissions were finally accepted. This issue is the second part of the special issue that contains 13 accepted articles.
This issue covers a variety of topics related to trustworthy recommendation and search, including five articles on the interpretability of recommendation and search, two articles on the security of recommendation and search (i.e., adversarial attacks and defenses), two papers on the robustness of search and click models, two articles on the robustness of recommendation models to the data sparsity, one article on the fairness of recommendation systems, and one article on the federated recommender system to protect user privacy. In what follows, we provide an overview of these accepted articles.
Interpretability is crucial for recommendation and search systems as it helps users to understand how results are generated, as well as being able to improve the trustworthiness of the system. Zhou et al. [
2022] argue that most existing sequence-based and graph-based methods learn user preferences in a black-box fashion, which leads to poor interpretability. In light of this, they propose an
explainable hyperbolic temporal point process for the user-item interaction sequence generation
(EHTPP), which introduces four influence factors from local and global perspectives. Notably, one of the influence factors, the popularity tendency factor, is used to interpret the recommendation results, as it indicates the occurrence probability of the user-item interaction. Leonhardt et al. [
2022] propose a two-stage SELECT-AND-RANK method for interpretable document ranking, which first extracts the given number of sentences from the input document from a denoised point of view and then only performs relevance estimation over these extracted sentences. Xue et al. [
2022] propose a novel sequential recommendation model that learns user representations from two distinct views, i.e., the item view and the factor view. Specifically, the item-view representation is learned from the sequence graph that includes the concrete timespans of the interaction between users and items, and the factor-view representation is learned from the sequence graph that is incorporated with the item genres by a hierarchical graph neural network, thereby providing explicit interpretability for recommendation results. Li et al. [
2022] propose a
Topic-aware Intention Network (TIN) for the explainable recommendation. It explains the recommendation results by a knowledge-enhanced neural topic model that learns the user intent. In this way, the generated user intent can be considered interpretable semantics. Quotations are crucial for successful explanations and persuasions. However, finding the right quotations to use can be challenging. Wang et al. [
2022b] propose a novel quotation recommendation framework for multi-party online conversations, which considers both semantic-level and topic-level information of conversations by jointly training semantic-based and topic-based recommendation models. Thus, the proposed method can well rank the candidate quotations by the similarities between conversation representations and quotations representations.
Adversarial attacks and defenses are essential in recommendation and search as they help identify vulnerabilities and improve the reliability of recommendation and search models against malicious attacks, which is crucial in preventing online spamming. Xu et al. [
2022] present a malicious attack on
graph learning-based collaborative filtering (GCLF), which uses a greedy strategy to search the local optimal perturbations and develops a proper attacking utility function to deal with the non-differentiable ranking-oriented metrics. In light of this security risk, a defense strategy based on measuring the suspicious score of each interaction and reducing the message weight of suspicious interactions during propagation is proposed to robustify GCLF further. Chen et al. [
2022] propose a
Word Substitution Ranking Attack on neural ranking models by adding adversarial perturbations to the target document, thereby promoting its ranking and exposure rate. Specifically, the proposed method is in the black-box attack setting where the model information is unavailable to the attackers. They present a
Pseudo Relevance-based ADversarial ranking Attack approach to finding the adversarial perturbations based on the generated gradients from a surrogate model.
The robustness of search and ranking models is a critical topic frequently discussed. One aspect of this is their ability to handle data noise and distributional shifts. He et al. [
2022] present an unsupervised semantic hashing framework, MASH, for similar text search. In addition, they introduce a relevance constraint and optimize hashing learning in the global space to ensure code balance. Notably, to ensure that the learned hash codes are robust to the noise, an improved SMASH model consisting of a noise-aware encoder-decoder framework is proposed to account for the noise degree of each text. On the other hand, Deffayet et al. [
2022] first analyze the limitations of the traditional offline evaluation protocol for click models; that is, they need to consider the robustness of these models to policy distributional shift. Then, the authors propose a new evaluation protocol that measures the robustness of six different click models under various shifts, training configurations, and downstream tasks. The study provides insights into factors that increase sensitivity to the policy distributional shift and offers guidelines to mitigate risks when deploying policies based on click models.
Data sparsity has been a significant challenge in recommender systems. Due to the effectiveness of the cross-domain information in alleviating the sparsity issue in recommender systems,
Cross-domain Recommendation (CR) task [Guo et al.
2022b,
2022c,
2023] is gaining immense attention. Guo et al. [
2022a] propose a novel end-to-end model,
DR-MTCDR, for the
multi-target cross-domain recommendation. The model aims to efficiently and effectively transfer trustworthy domain-shared information across multiple domains while avoiding adverse transfer problems. To achieve this, DR-MTCDR uses a unified module to capture disentangled domain-shared and domain-specific information, which supports various domains’ recommendations and is insensitive to the number of domains. In addition, we have an accepted article addressing the data sparsity issue in news recommendations. To address the data sparsity, Meng et al. [
2022] integrate user-news interaction relationships with the news click sequences to build a global heterogeneous graph. Then, they propose a
heterogeneous transition attention networks-based news recommendation framework, GAINRec, which can learn the news transition patterns and capture the common behavior patterns of users over the constructed global heterogeneous graph. Furthermore, the edge attention weights could be visualized to enhance the interpretability and reliability of the proposed method.
With the growing concerns about privacy and the new privacy protection regulations (e.g., GDPR in the EU and CCPA in the USA),
federated recommenders (FedRecs) have recently emerged as a privacy-preserving solution to collaboratively learn a recommendation model among personal devices without uploading user data to a central server. Nevertheless, performing complex operations, e.g., similarity search via inner product, on resource-constrained devices remains a significant challenge for current federated recommender systems. Zhang et al. [
2022a] propose a lightweight matrix factorization-based federated recommendation framework, called
LightFR, which employs the hash technique to generate high-quality binary codes to perform faster online inference and reduce memory consumption.
Fairness is another important dimension in trustworthy recommendation systems. Fairness means that the recommendations provided by the system should be unbiased and not discriminate. Zhao et al. [
2022] propose an adaptive and fair task recommendation framework in spatial crowdsourcing [Nguyen et al.
2018]. Concretely, it first learns workers’ travel-intention-based preferences by capturing both workers’ hometown and out-of-town preferences adaptively based on their locations, and then introduces a fair top-
k recommendation method.