Item

Abstract

Our increasing reliance on complex algorithms for recommendations calls for models and methods for explainable, scrutable, and trustworthy AI. While explainability is required for understanding the relationships between model inputs and outputs, a scrutable system allows us to modify its behavior as desired. These properties help bridge the gap between our expectations and the algorithm’s behavior and accordingly boost our trust in AI. Aiming to cope with information overload, recommender systems play a crucial role in filtering content (such as products, news, songs, and movies) and shaping a personalized experience for their users. Consequently, there has been a growing demand from the information consumers to receive proper explanations for their personalized recommendations. These explanations aim at helping users understand why certain items are recommended to them and how their previous inputs to the system relate to the generation of such recommendations. Besides, in the event of receiving undesirable content, explanations could possibly contain valuable information as to how the system’s behavior can be modified accordingly. In this thesis, we present our contributions towards explainability and scrutability of recommender systems: • We introduce a user-centric framework, FAIRY, for discovering and ranking post-hoc explanations for the social feeds generated by black-box platforms. These explanations reveal relationships between users’ profiles and their feed items and are extracted from the local interaction graphs of users. FAIRY employs a learning-to-rank (LTR) method to score candidate explanations based on their relevance and surprisal. • We propose a method, PRINCE, to facilitate provider-side explainability in graph-based recommender systems that use personalized PageRank at their core. PRINCE explanations are comprehensible for users, because they present subsets of the user’s prior actions responsible for the received recommendations. PRINCE operates in a counterfactual setup and builds on a polynomial-time algorithm for finding the smallest counterfactual explanations. • We propose a human-in-the-loop framework, ELIXIR, for enhancing scrutability and subsequently the recommendation models by leveraging user feedback on explanations. ELIXIR enables recommender systems to collect user feedback on pairs of recommendations and explanations. The feedback is incorporated into the model by imposing a soft constraint for learning user-specific item representations. We evaluate all proposed models and methods with real user studies and demonstrate their benefits at achieving explainability and scrutability in recommender systems.