A NeuMF-Based Framework for Sparse E-Commerce Recommendation with Cold-Start Fallback

Abstract

Personalized recommender systems have become essential in modern e-commerce by improving product discovery and enhancing user experience. However, recommendation performance is often limited by sparse user–item interactions and cold-start scenarios, where insufficient historical data reduces personalization accuracy. This study evaluates a hybrid recommendation framework that uses a NeuMF-based model for users with historical interactions and a popularity-based fallback for completely unseen users. The NeuMF model combines Generalized Matrix Factorization (GMF) and Multi-Layer Perceptron (MLP) components, with Batch Normalization and Dropout examined through controlled ablation experiments. The evaluation separates three tasks: rating prediction, full-catalog Top-K ranking, and true user cold-start recommendation. On the Amazon Electronics dataset, the proposed NeuMF variant achieved a lower MAE than the baseline NeuMF model (0.8603 vs. 0.8837), while the baseline obtained a lower RMSE (1.2172 vs. 1.2362). The difference in NDCG@10 between the proposed and baseline models was not statistically significant. In full-catalog ranking, LightGCN achieved the strongest overall performance among the evaluated models. For completely unseen users, where no learned user embedding was available, a popularity-based fallback provided better ranking performance than the other tested non-personalized strategies. These results show that recommendation performance depends on the evaluation task and that rating prediction, Top-K ranking, and true cold-start recommendation should be assessed separately.