CoFARS-Sparse: A Context-Aware Recommendation Framework for Sparse E-Commerce Data

Abstract

Predicting user behavior in extremely sparse data environments where users interact only once remains a fundamental challenge in e-commerce recommendation. While specialized domains like home decor face severe sparsity (where 87% of users exhibit single-interaction cold-start behavior), conventional context-aware sequential models suffer sharp performance degradation when interaction sequences lack length. To address this, we introduce a regional home-product e-commerce dataset reflecting extreme sparsity alongside CoFARS-Sparse, a tailored recommendation framework. Our approach replaces opaque embeddings with an interpretable Probability Encoder over product attributes, uses a Static Context Matcher to bypass unreliable temporal graph construction, and employs a Hybrid User Encoder that dynamically adapts its routing logic based on user history length. Experiments demonstrate that CoFARS-Sparse achieves an AUC of 0.9165 and an NDCG@5 of 0.2826 under a 1-vs-99 ranking protocol, outperforming state-of-the-art sequential and graph baselines while significantly reducing inference latency.