Towards Trustworthy Recommender System: A Faithful and Responsible Recommendation Perspective

Recommender systems (RecSys) become increasingly prevalent in modern society, offering personalized information filtering to alleviate information overload and significantly impacting various human online activities. Machine learning-based recommendation methods have been extensively developed in recent years to achieve more accurate recommendations, with some of these approaches having been extensively deployed in industrial applications, such as the Deep Interest Network (DIN). Despite their widespread use, researchers and practitioners have highlighted various trustworthiness issues inherent in these systems, including bias and promoting polarization issues. In order to better serve users and comply with regulations pertaining to recommendation algorithms established by different countries, it is essential to consider the trustworthiness issues of recommender systems.

This research focuses on trustworthiness in recommendation from two perspectives of user-centered principles: faithfulness and responsibility. On the one hand, collected recommendation data may not faithfully reflect user preferences, especially those of the service stage, due to bias[2, 3] and temporal effects,[4,5]etc. Achieving faithful recommendations with such data is crucial to ensure user satisfaction, i.e., making recommendations faithfully reflect user preferences during the testing. On the other hand, recommender systems could not only cater to user preferences [1] but also unconsciously and unintentionally affect (or even manipulate) user preferences. In the recommendation process, controlling the influence of recommender systems, such as avoiding potential opinion polarization, to provide responsible recommendations is also an important aspect of building trustworthy recommender systems. Consequently, there raise four research questions on the two aspects:

Our objective is to achieve faithful and responsible recommendations for users while addressing these research questions. We attribute unfaithful recommendation to the discrepancies between the training data and the service objectives, which we formulate as different data shift problems (RQ1 and RQ2). We provide systematic analyses for these data shift problems from causal perspectives and develop several causality-inspired solutions to enhance recommendation faithfulness. In pursuit of responsible recommendations, we investigate the effect of recommender systems on users from a causal perspective. We develop a causal effect evaluation and adjustment framework to quantify and control the influence of recommender systems on user preferences (RQ3 and RQ4).