DualDS: A dual discriminative rating elicitation framework for cold start recommendation

Cold start problem is challenging because no prior knowledge can be used in recommendation. To address this cold start scenario, rating elicitation is usually employed, which profiles cold user or item by acquiring ratings during an initial interview. However, how to elicit the most valuable ratings is still an open problem. Intuitively, category labels which indicate user preferences and item attributes are quite useful. For example, category information can be served as a guidance to generate a set of queries which can largely capture the interests of cold users, and thus appealing recommendation lists are more likely to be returned. Therefore, we exploit category labels as supervised information to select discriminative queries. Furthermore, by exploring the correlation between users and items, a dual regularization is developed to jointly select optimal representatives. As a consequent, a novel Dual Discriminative Selection (DualDS) framework for rating elicitation is proposed in this paper, by integrating discriminative selection with dual regularization. Experiments on two real-world datasets demonstrate the effectiveness of DualDS for cold start recommendation.