Graph contrastive learning for recommendation with generative data augmentation

Xiaoge Li

doi:10.1007/s00530-024-01375-z

2024-06-06

PRE

DOI：10.1007/s00530-024-01375-z

原文链接

原文求助

摘要

En 中文

Graph Neural Networks (GNNs) have been successfully adopted in recommender systems by virtue of the message-passing that implicitly captures collaborative effect. However, in practical recommendation scenarios, user-item interaction data is often sparse and exhibits a skewed distribution. To address these issues, some contrastive learning methods based on data augmentation are applied to recommender systems to enhance the representation of users and items. Nevertheless, many data enhancements solely rely on graph topology, missing crucial structural information and potentially biasing the model. In this paper, we propose a contrastive learning recommendation framework called GDA-GCL based on a generative model data augmentation strategy. Specifically, we use the Conditional Variational Autoencoder(CVAE) generative model to learn the distribution of neighbor node features conditioned on the features of the central node. Due to the randomness of resampling, we design a mirror graph comparison strategy to generate different comparison views, which introduces additional high-quality training signals into the GNN paradigm. Experimental results on three real-world public datasets demonstrate that GDA-GCL achieves significant improvement in performance over various baseline methods. Extensive analysis, including ablation studies, has demonstrated the effectiveness and robustness of our proposed generative data-augmented contrastive recommendation framework in addressing the data sparsity issue in recommendation systems.

Keyword:

Collaborative filtering

Recommendation systems

Graph neural networks

Data augmentation

Generative models