Quantifying and Mitigating Popularity Bias in Conversational Recommender Systems

Abstract

Conversational recommender systems (CRS) have shown great success in\naccurately capturing a user's current and detailed preference through the\nmulti-round interaction cycle while effectively guiding users to a more\npersonalized recommendation. Perhaps surprisingly, conversational recommender\nsystems can be plagued by popularity bias, much like traditional recommender\nsystems. In this paper, we systematically study the problem of popularity bias\nin CRSs. We demonstrate the existence of popularity bias in existing\nstate-of-the-art CRSs from an exposure rate, a success rate, and a\nconversational utility perspective, and propose a suite of popularity bias\nmetrics designed specifically for the CRS setting. We then introduce a\ndebiasing framework with three unique features: (i) Popularity-Aware Focused\nLearning to reduce the popularity-distorting impact on preference prediction;\n(ii) Cold-Start Item Embedding Reconstruction via Attribute Mapping, to improve\nthe modeling of cold-start items; and (iii) Dual-Policy Learning, to better\nguide the CRS when dealing with either popular or unpopular items. Through\nextensive experiments on two frequently used CRS datasets, we find the proposed\nmodel-agnostic debiasing framework not only mitigates the popularity bias in\nstate-of-the-art CRSs but also improves the overall recommendation performance.\n