Online Distillation and Preferences Fusion for Graph Convolutional Network-Based Sequential Recommendation

Sequential recommendations make an attempt to predict the next item that a user will interact with based on their historical behavior sequence. Recently, considering the relationship learning ability of graph convolutional network (GCN), a number of GCN-based sequence recommendation models have emerged. However, in real-world applications, sparse interactions are common, with early and current short-term preferences playing diverse roles in sequential recommendation. As a result, vanilla GCNs fail to investigate the explicit relationship between these early and current short-term preferences. To address the above limitations, we propose a scheme of Online Distillation and Preferences Fusion for GCN-based sequential recommendation (ODPF). Specifically, our approach performs online distillation among multiple networks to learn item feature representations. To distinguish between early and recent short-term preferences, we divide each sequence into two subsequences and construct two graphs separately. On this basis, a fusion network is introduced to capture more accurate preferences by fusing these two types of preferences. Experimental evaluations conducted on two public datasets demonstrate that our proposed method outperforms recent state-of-the-art methods in terms of recommendation precision.