Modeling Tradeoffs Using Preference-Based Feedback in Session-Based Recommender Systems

Abstract

Cold-start is a challenge in most data-driven approaches across various paradigms of recommender systems research. Knowledge integration is becoming essential to alleviate the problem. Despite the availability of knowledge representations like the knowledge graph, there is a need for a richer semantic model to capture relationships among products in a domain. Recommender systems inspired by case-based reasoning uses rich domain knowledge in their recommendation process. Conversational case-based reasoning recommender systems (CCBR-RS) are session-based recommender systems that operate under cold-start assumptions. User preferences in such scenarios can only be learned based on the context of the current session. Tradeoffs have been used in literature to model the relationship among products in recommending appropriate products. The models in earlier works have been based on crisp sets, which lack flexibility in capturing varying degrees of tradeoffs which is a more natural way of dealing with tradeoffs. This work proposes a flexible tradeoff representation scheme based on fuzzy sets to model tradeoffs in terms of linguistic terms and a fuzzy inference system that measures the similarity among tradeoffs represented in linguistic terms. We demonstrate the performance improvement in utilizing a more flexible model of tradeoffs by evaluating the methods on three datasets in a CCBR-RS framework.