1 Introduction
Recommendation and search systems have already become indispensable means for 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 assistant, 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. Firstly, 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]. Secondly, 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; Yin et al.
2012]; while security and privacy respectively emphasize a model’s ability to handle cyber-attacks [Zhang et al.
2021b;,
2022] and to prevent personal information leakage [Zhang and Yin
2022; Zhang et al.
2021c;,
2021a; Yuan et al.
2023; Wang et al.
2022b]. 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. In Part 1 of this issue, we present 10 accepted papers and leave the remaining accepted papers to the upcoming Part 2. This issue covers a variety of topics related to trustworthy recommendation and search, including four papers on fairness/bias issues in recommendation and ranking, three papers on adversarial attacks and/or their countermeasures for recommender systems, two papers on privacy-aware recommendation, and one paper on the interpretability of recommender systems. In what follows, we provide an overview on these accepted articles.
Ensuring fairness and mitigating algorithmic bias are attracting much attention within the recommendation and search community. By providing a comprehensive survey, Wang et al. [
2022a] have summarized various fairness challenges in the recommendation context, reviewed key recommendation datasets and measurements in fairness studies, and provided a well-structured taxonomy of fairness-aware recommendation methods. In this survey, an insightful list of suggestions on future research on the recommendation fairness is also provided, which will benefit subsequent studies in this area. Meanwhile, three specific biases are respectively investigated in the other three papers. He et al. [
2022] formulate and address the problem of items’ confounding features from the causal perspective, where the proposed solution performs intervened inference with do-calculus. In addition, the authors further propose a mixture-of-experts model architecture that models each value of confounding feature with a separate expert module, so as to significantly lower the computational complexity for evaluating do-calculus. The designed solution thus represents a strong synergy between model expressiveness (i.e., accuracy) and efficiency for casual recommendation. In Liu et al. [
2022], the authors study the popularity bias in recommender systems, and unlike most existing studies that only consider fairness at either the user side or item side, their proposed solution jointly mitigates the popularity bias for both users and items for recommendation. Besides, the solution is able to dynamically adapt to different input users and items to handle the differences in their popularity bias, thus contributing to an outstanding debiasing efficacy. In the context of
learning-to-rank (LTR) algorithms, Oosterhuis [
2022] presents a novel solution to the long-standing position bias issues via the lens of counterfactual doubly-robust estimation. As the first doubly-robust solution to position bias in LTR, Oosterhuis proposes to counteract the unobservable treatment (i.e., user examination) by using the expected treatment per rank instead of the actual treatment. The designed estimator is more robust and performant than existing approaches based on inverse propensity scoring, and it further provides the most robust theoretical guarantees of all known LTR estimators.
The need for privacy protection is surging in various web services, and it is no exception for recommender systems as they heavily rely on the sensitive user data to facilitate personalized recommendation. On the one hand, from the perspective of service providers, Chen et al. [
2022] present a new take on the common practice of recommendation platforms – offering users binary choice on data disclosure. By designing a privacy-aware recommendation framework that gives users fine-grained control over their data, the authors perform comprehensive experiments in a simulated real-world environment to uncover how different privacy levels impact users’ information disclosure willingness and the platform’s revenue. An important message from this paper is that privacy mechanisms with finer split granularity and more unrestrained disclosure strategy can bring better results for both consumers and platforms than the “all or nothing” mechanism adopted by most real-world applications. On the other hand, from the perspective of users, Xin et al. [
2022] investigate the problem of user behavior leakage in recommender systems. The authors show that the sensitive historical interaction of a user can be inferred from the currently observed system exposure for this user, a.k.a. membership inference. As a remedy, a privacy protection mechanism is proposed to perturb a subset of exposed items. Because the wide access to exposure data put users’ interaction history at a high risk of leakage, this paper will open up an important research topic in privacy-aware recommendation.
Another important topic often discussed in conjunction with privacy in recommendation and search is the adversarial attacks and their countermeasures. In the context of hashing retrieval, given that most existing adversarial attack models assume the impractical white-box setting and are inefficient to train, Zhu et al. [
2022] propose an efficient black-box attack model against deep cross-modal hashing retrieval. Specifically, the solution proposes a multi-modal knockoff-driven adversarial generation framework to achieve efficient adversarial example generation. This allows the attacker to efficiently generate quality adversarial examples by forward-propagation with only given benign images under the black-box setting. Meanwhile, Nguyen et al. [
2022] have studied poisoning attack on state-of-the-art recommenders based on
graph neural networks (GNNs). Attacking GNN-based recommenders is more challenging than attacking a plain GNN due to the heterogeneity of network structure and the entanglement between users and items. As such, the authors propose to surrogate a recommendation model, as well as to generate fake users and user-item interactions while preserving the user-item correlations for recommendation accuracy. The proposed solution is also resistant to various protection mechanisms, shedding light on future research on attack-resistant recommender systems. As a possible countermeasure to such attacks, Ye et al. [
2022] propose to improve the robustness of GNN-based recommenders by jointly denoising the structure space and perturbing the embedding space. Notably, in the embedding space, an in-distribution perturbation method is designed to simulate adversarial attacks, thus providing a boost in the recommender’s robustness to noisy interactions in the training data.
We also have an accepted article focusing on explainable recommendation. Motivated by the fact that exploring users’ fine-grained preferences as well as the relationships among those preferences could improve the recommendation performance, Dong et al. [
2022] propose a dual preference distribution learning framework to jointly learn a user’s general preference to items and the user’s specific preference to item attributes. To support interpretability, a preferred attribute profile is summarized for each user, where the explanation for each recommended item can be generated by checking the overlap between its own and the user’s preferred attributes.
3 Conclusions
In summary, these accepted articles are a strong reflection of the depth and breadth of current research on trustworthy recommendation and search. A wide range of research questions have been discussed in Part 1 of this special issue, including how to achieve a sensible privacy-utility trade-off, how to mitigate algorithmic bias, how to understand and prevent possible adversarial attacks, and how to preserve explainability of a complex algorithm. In the meantime, there are still open challenges to be addressed in this research sector, such as a recommendation/retrieval model’s awareness to data veracity, capability of performing reasoning when interacting with users, and ability to handle high throughput data streams. In Part 2 of the special issue, we will introduce more work in those areas.