Dual Learning for Explainable Recommendation: Towards Unifying User Preference Prediction and Review Generation

In many recommender systems, users express item opinions through two kinds of behaviors: giving preferences and writing detailed reviews. As both kinds of behaviors reflect users' assessment of items, review enhanced recommender systems leverage these two kinds of user behaviors to boost recommendation performance. On the one hand, researchers proposed to better model the user and item embeddings with additional review information for enhancing preference prediction accuracy. On the other hand, some recent works focused on automatically generating item reviews for recommendation explanations with related user and item embeddings. We argue that, while the task of preference prediction with the accuracy goal is well recognized in the community, the task of generating reviews for explainable recommendation is also important to gain user trust and increase conversion rate. Some preliminary attempts have considered jointly modeling these two tasks, with the user and item embeddings are shared. These studies empirically showed that these two tasks are correlated, and jointly modeling them would benefit the performance of both tasks. In this paper, we make a further study of unifying these two tasks for explainable recommendation. Instead of simply correlating these two tasks with shared user and item embeddings, we argue that these two tasks are presented in dual forms. In other words, the input of the primal preference prediction task p(R vertical bar C) is exactly the output of the dual review generation task p(C vertical bar R), with R and C denote the preference value space and review space. Therefore, we could explicitly model the probabilistic correlation between these two dual tasks with p(R, C)= p(R vertical bar C)p(C)= p(C vertical bar R)p(R). We design a unified dual framework of how to inject the probabilistic duality of the two tasks in the training stage. Furthermore, as the detailed preference and review information are not available for each user-item pair in the test stage, we propose a transfer learning based model for preference prediction and review generation. Finally, extensive experimental results on two real-world datasets clearly show the effectiveness of our proposed model for both user preference prediction and review generation.