Dual representation propagation comparative learning for recommendations

Abstract

Graph neural networks combined with comparative learning have become a very popular paradigm in recommender systems. However, most methods still suffer from data sparsity, noise and difficulty in extracting multi-granularity information. To address these limitations, we propose a Dual Representation Propagation Comparative Learning(DRPCL) method, which uses propagation representations based on two rules to extract multi-granularity signals, including graph convolutional propagation pathway and message node propagation pathway. Where the message node propagation pathway generates message nodes containing local information, and the message nodes are used as the hub to extract global signals, so as to fuse local and global signals. And the dual-pathway node representations generate two comparative views for misaligned comparative learning to alleviate the problem of sparse data. At the same time, denoising auxiliary supervision signals are generated to affect the propagation of node representations to reduce the negative effects of noise. Experimental results show that our DRPCL is able to demonstrate performance superiority over other bases on different datasets. Some in-depth experimental analysis demonstrates the robustness of DRPCL against data sparsity and noise.