SEMA: Deeply Learning Semantic Meanings and Temporal Dynamics for Recommendations

Abstract

Personalization plays an essential role in recommender systems, in which the key task is to predict the personalized rating of users on new items. Recommender systems usually apply collaborative filtering techniques to make rating prediction. In recent years, some studies pay attention on learning semantic meanings from textual content of items or temporal dynamics from historical information of users in order to improve rating prediction. However, these studies often apply shallow or flat modeling methods and model users and items in an asymmetrical manner; the improvement is considerably limited. In this paper, we propose a new recommendation framework called SEMA to deeply learn ${S}$ emantic $\text{m}{E}$ anings and te ${M}$ poral dyn ${A}$ mics by developing hierarchical and symmetrical recurrent neural networks (RNNs). Our SEMA has three important characteristics: 1) deep learning-based: SEMA leverages deep learning-based models to capture semantic meanings from textual content and temporal dynamics from historical information rather than applying shallow methods, e.g., the bag-of-words method for textual content and the decay method for temporal dynamics; 2) hierarchical: SEMA learns both semantic meanings and temporal dynamics in a unified hierarchical RNN to mutually reinforce each other, instead of combining them flatly; and 3) symmetrical: SEMA symmetrically builds two hierarchical RNNs for users and items to model their own semantic meanings and temporal dynamics, because users and items are essentially dual in recommender systems. We conduct a comprehensive performance evaluation for SEMA using two large-scale real-world review data sets collected from Amazon and Yelp. Experimental results show that SEMA achieves significantly superior recommendation quality compared with other state-of-the-art recommendation techniques.