Bilateral Unsymmetrical Graph Contrastive Learning for Recommendation

TL;DR

A novel framework for recommendation tasks called Bilateral Unsymmetrical Graph Contrastive Learning (BusGCL) that considers the bilateral unsymmetry on user-item node relation density for sliced user and item graph reasoning better with bilateral slicing contrastive training is proposed.

Abstract

Recent methods utilize graph contrastive Learning within graph-structured user-item interaction data for collaborative filtering and have demonstrated their efficacy in recommendation tasks. However, they ignore that the difference relation density of nodes between the user- and item-side causes the adaptability of graphs on bilateral nodes to be different after multi-hop graph interaction calculation, which limits existing models to achieve ideal results. To solve this issue, we propose a novel framework for recommendation tasks called Bilateral Unsymmetrical Graph Contrastive Learning (BusGCL) that consider the bilateral unsymmetry on user-item node relation density for sliced user and item graph reasoning better with bilateral slicing contrastive training. Especially, taking into account the aggregation ability of hypergraph-based graph convolutional network (GCN) in digging implicit similarities is more suitable for user nodes, embeddings generated from three different modules: hypergraph-based GCN, GCN and perturbed GCN, are sliced into two subviews by the user- and item-side respectively, and selectively combined into subview pairs bilaterally based on the characteristics of inter-node relation structure. Furthermore, to align the distribution of user and item embeddings after aggregation, a dispersing loss is leveraged to adjust the mutual distance between all embeddings for maintaining learning ability. Comprehensive experiments on two public datasets have proved the superiority of BusGCL in comparison to various recommendation methods. Other models can simply utilize our bilateral slicing contrastive learning to enhance recommending performance without incurring extra expenses.

Figures (5)

• Figure 1: Illustration of difference in the relation density between bilateral nodes. (a) means that user nodes often have denser inter-node relationships than item nodes. (b) visualizes the different distribution of embeddings generated from a 1-layer LightGCN LightGCN-He2020LightGCNSA on Yelp dataset by t-SNE. (c) counts the number of 2-hop neighbours, by normalizing them into a same total number.
• Figure 2: Illustration of BusGCL framework. An adjacency matrix which represents user-item interaction graph, passes through a Multi-structurally Graph Model that contains three variants of GCNs to form three embedding matrices. And three embedding matrices are sliced by both user-side and item-side for bilateral slicing contrastive learning before recommendation predictions.
• Figure 3: The effect of GCN layers, the data of the two images were measured on the Yelp and Last.FM datasets, respectively.
• Figure 4: Influence of weight and temperature coefficient about dispersing loss on recommendation performance on Yelp dataset, that measurement by Recall@20 is presented in the left image and NDCG@20 on the right.
• Figure 5: Visualization of distribution from different stages of BusGCL training by t-SNE on Yelp dataset. (a) shows embeddings after GCN process and readout directly. (b) is after bilateral subview CL. (c) adds dispersing loss. Additionally, (d) and (e) illustrate the situation when subviews on both sides are selected from hyperGCN/GCN with perturbing.