Explainable Recommendation via Neural Rating Regression and Fine-Grained Sentiment Perception

Compared with traditional recommendation systems, explainable recommendation systems have certain advantages in terms of system transparency, result credibility, and user satisfaction. However, the existing text explanation generation methods are often limited by pre-defined templates which limit the ability of text expression. The free-style text generation methods have stronger expressive ability, but are less controllable and ignore fine-grained sentiment perception from user comments. In this paper, a Dual Learning-based Explainable Recommendation Model (called DLER) is proposed, which uses a dual learning mechanism to perform rating prediction and explanation generation respectively. The parameters can be adjusted iteratively via the collaborative promotion between the two phases. A rating prediction algorithm based on neural rating regression and an explanation generation algorithm based on fine-grained sentiment perception are respectively proposed. On the one hand, the ratings are predicted via MLP. On the other hand, users' fine-grained sentiments are perceived by analyzing comments, which will be used for GRU-based explanation generation. The experiments demonstrate the effectiveness and the efficiency of our proposed method in comparison with traditional methods.