A key challenge in recommender systems is how to provide recommendations for cold-start users about which the system has no prior knowledge. A common approach to this problem is to conduct a brief interview with the user to elicit their preferences on a number of informative entities. While most proposed approaches have focused on eliciting item-specific preferences from users, users may be able to better opine on broader and more descriptive properties of items, denoted as descriptive entities.
While this focus on items, denoted recommendable entities, can be attributed largely to the lack of available and appropriate datasets, the recently published MindReader dataset alleviates this issue.
In this work, we perform a comprehensive study of interviewing strategies and models of recommendation including state-of-the-art methods, and evaluate the effectiveness of allowing interviewing systems to ask towards descriptive entities in the MindReader dataset, which we further extend to 1,736 users and 174,872 ratings.
In order to construct optimal interviews, we propose a novel, adaptive interview learning approach, as well as approaches based on deep reinforcement learning. For making recommendations from user answers, we further propose a linear combination of Personalised PageRank which learns the importance of knowledge- and collaborative graphs through a pairwise ranking loss. Our findings show that almost all models can improve performance with broader questions, allowing the interview to be shortened by ~4 questions when asking towards descriptive rather than recommendable entities. Our findings show that especially knowledge-aware approaches can benefit greatly from descriptive entity preferences in cold-start interviews and outperform state-of-the-art methods in both recommendation quality and diversity.