Co-Attentive Graph Learning for Session-based Recommendation
Fei Cai, Zhiqiang Pan · 2022 5th International Conference on Data Science and Information Technology (DSIT) · 2022
The goal of session-based recommendation is to predict user's action at the next timestamp by modeling his recent limited behaviors. Existing methods mainly concentrate on the sequential signal or pairwise transition relation between items using recurrent neural networks (RNN) or graph neural networks (GNN). However, the sequential methods fail to consider the transition relation between items while the GNN-based models face a serious over-smoothing problem. Thus, we propose a Co-Attentive Graph Learning (CAGL) method for session-based recommendation, which introduces a co-attention network to prevent over-smoothing, thus learning accurate representations for items. Extensive experiments are conducted on two public datasets, i.e., Diginetica and Gowalla. The results show that CAGL can beat the state-of-the-art baselines in terms of Recall and MRR.