Enhancing e-commerce representation learning via hypergraph contrastive learning and interpretable LLM-driven analysis

E-commerce has experienced significant growth recently, generating vast amounts of data on user preferences, interactions, and purchase patterns. Effectively modeling and representing users and products in these online ecosystems is crucial for various applications. However, existing approaches for e-commerce representation learning face several limitations: (i) they primarily consider user behavior patterns while ignoring rich group-wise relationships; (ii) some works focus on either user or product representations, failing to learn both simultaneously; (iii) results on downstream tasks are generated by "black box" models, making it difficult to interpret prediction results. To address these challenges, we propose RelationExpert, a general e-commerce representation learning framework. It consists of two components: RelationEmbed, an ecommerce representation learning foundation model, and TaskReport, an interpretability-driven LLM module. RelationEmbed is a self-supervised hypergraph contrastive learning model to capture multi-modal features and rich group-wise relationships among unlabeled data, i.e., merchants, customers, and products. TaskReport generates interpretable reports that explain the results of downstream tasks utilizing RelationEmbed’s learned embeddings. As a result, (i) General: RelationExpert is applicable to various e-commerce related tasks; (ii) Novel and Powerful: as the first e-commerce hypergraph contrastive learning framework, RelationEmbed significantly outperforms existing methods across eight downstream tasks on two markets; (iii) Interpretable and Reliable: TaskReport provides clear insights into "black box" results and delivers reliable reports with high factuality and clarity.