Dual-Channel Convolutional Recurrent Networks for Session-Based Recommendation

Abstract

Recommender systems assist a Web application user in satisfying their needs or interests based on the user profile and past activities. Yet due to privacy and other concerns, some applications and services only keep anonymous information. A session-based recommender system (SRS) predicts the next item by exploring only anonymous user-item behavior orders during ongoing sessions. Recurrent neural networks (RNNs) and their two variants have dominated the research on SRS. However, there are two shortcomings in these RNN-based methods: (1) RNNs easily generate false dependencies because RNNs assume all adjacent items are highly dependent on each other; (2) the sequentially connected architecture of RNNs can only capture the point-level dependencies but ignoring neglecting the union-level dependencies. This paper proposes a Dual-channel Convolutional Recurrent Neural Network (D-CRNN) model to address these problems. This hybrid model leverages RNN to explore complex long-term dependencies and combines CNN to extract the union-level context features, which help to reduce the noise. The hybrid model was evaluated on three commonly used real-world datasets. The experimental results on Diginetica dataset D-CRNN showed an improvement of 5.8% and 4.8% respectively in terms of Recall@10 and MRR@10, demonstrating the effectiveness of D-CRNN on the session-based recommendation.