Dual-view data augmentation at subgraph level and graph contrastive learning for sequential recommendation

Caihong Mu, Haikun Yu, Lang Qin, Yi Liu

Abstract: The existing sequential recommendation algorithms generally face the problems of not fully utilizing item relationships across sequences, being incapable of effectively capturing users’ global preferences, and being susceptible to data sparsity. To address above problems, we propose a dual-view data augmentation at subgraph level and graph contrastive learning for sequential recommendation (DSGCL). Firstly, we construct a weighted sequential transition global graph and an item correlation global graph based on sequential interaction data to utilize item relationships across sequences. Secondly, in order to mitigate the data sparsity problem, we construct a pair of augmented subgraphs for each global graph by using data augmentation at subgraph level, and capture user’s global preferences by using graph neural networks on the augmented subgraphs. Consistency between the same user preference learnt from different augmented subgraphs is ensured by employing the graph contrastive learning method, which also helps to better distinguish the difference in preferences between users. Experimental results on multiple sequential recommendation datasets show that the DSGCL has a better performance compared to other advanced methods. Besides, ablation experiments validate the effectiveness of each module in the model.