E-Commerce User Shopping Preference Ranking Toward Million-Scale Products: A Hierarchical Feature Learning and Huge Purchase Graph Clustering Framework

Modern e-commerce platforms face significant challenges in understanding user preferences at scale, particularly when dealing with million-product catalogs and diverse shopping behaviors. Traditional approaches struggle to (1) extract meaningful patterns from sparse, high-dimensional behavioral data, (2) efficiently cluster users with similar preferences, and (3) generate personalized recommendations that balance individual tastes with broader community trends. We propose a scalable graph-based framework that addresses these challenges through three key innovations. First, our hierarchical feature learning system processes diverse behavioral signals - including preferred product categories, dress colors, and visual styles - using weakly supervised learning to ensure comprehensive preference coverage. Second, a geometric selection strategy filters noisy features, producing compact yet expressive user representations that are projected into a latent topic space for precise differentiation. Third, we construct a weighted user similarity graph enabling efficient community detection at scale, revealing preference patterns ranging from bold aesthetics to minimalist styles. The complete pipeline integrates these components with a ranking module that fuses individual user profiles with community embeddings, achieving personalized recommendations that reflect both specific user behaviors and broader preference trends. Extensive evaluations on a dataset of over one million users demonstrate our framework’s superiority in both clustering quality (28% higher NMI than baseline methods) and recommendation accuracy (15% improvement in recall@k), while maintaining computational efficiency suitable for production deployment.