A Comparative Study of Compound Critique Generation in Conversational Recommender Systems

Critiquing techniques provide an easy way for users to feedback their preferences over one or several attributes of the products in a conversational recommender system. While unit critiques only allow users to critique one attribute of the products each time, a well-generated set of compound critiques enables users to input their preferences on several attributes at the same time, and can potentially shorten the interaction cycles in finding the target products. As a result, the dynamic generation of compound critiques is a critical issue for designing the critique-based conversational recommender systems. In earlier research the Apriori algorithm has been adopted to generate compound critiques from the given data set. In this paper we propose an alternative approach for generating compound critiques based on the multi-attribute utility theory (MAUT). Our approach automatically updates the weights of the product attributes as the result of the interactive critiquing process. This modification of weights is then used to determine the compound critiques according to those products with the highest utility values. Our experiments show that the compound critiques generated by this approach are more efficient in helping users find their target products than those generated by the Apriori algorithm.