Multi-Graph Contrastive Learning With LLM-Empowered Semantic Augmentation for E-Commerce Recommendation

Recommender systems are crucial in improving the user shopping experience and business revenue in e-commerce. Graph neural networks (GNNs) become a prominent method for their powerful high-order interaction capabilities. Nevertheless, it has been a long-standing challenge for GNN-based methods to learn sufficient feature representations for user or product nodes with sparse interaction information. To tackle this challenge, researchers introduce side information to enhance the learning of feature representations. However, they fail to fully exploit the potential valuable complementary relationships between products and neglect the semantic gap between user reviews. To leverage the information effectively, we propose Multi-Graph Contrastive Learning with LLM-empowered Semantic Augmentation for E-commerce Recommendation (MCLR). First, we leverage Large Language Models (LLMs) to explore the complementary relationship between products and standardize the user reviews to bridge the semantic gap. Then, we construct a user-product interaction graph based on historical interactions, a product-product complementarity graph based on the complementary relationship, and a user-user similarity graph based on standardized reviews to capture different associations between entities from multiple views. Moreover, we design a tailored multi-graph contrastive learning mechanism to fuse the information from the three graphs, thereby mitigating insufficient feature representation learning for users and products. Finally, extensive experiments are conducted to validate the superiority of the proposed method.