Variational Reasoning about User Preferences for Conversational Recommendation

Conversational recommender systems (CRSs) provide recommendations through interactive conversations. CRSs typically provide recommendations through relatively straightforward interactions, where the system continuously inquires about a user's explicit attribute-aware preferences and then decides which items to recommend. In addition, topic tracking is often used to provide naturally sounding responses. However, merely tracking topics is not enough to recognize a user's real preferences in a dialogue.

In this paper, we address the problem of accurately recognizing and maintaining user preferences in CRSs. Three challenges come with this problem: (1) An ongoing dialogue only provides the user's short-term feedback; (2) Annotations of user preferences are not available; and (3) There may be complex semantic correlations among items that feature in a dialogue. We tackle these challenges by proposing an end-to-end variational reasoning approach to the task of conversational recommendation. We model both long-term preferences and short-term preferences as latent variables with topical priors for explicit long-term and short-term preference exploration, respectively. We use an efficient stochastic gradient variational Bayesian (SGVB) estimator for optimizing the derived evidence lower bound. A policy network is then used to predict topics for a clarification utterance or items for a recommendation response. The use of explicit sequences of preferences with multi-hop reasoning in a heterogeneous knowledge graph helps to provide more accurate conversational recommendation results.

Extensive experiments conducted on two benchmark datasets show that our proposed method outperforms state-of-the-art baselines in terms of both objective and subjective evaluation metrics.