PERFORMANCE EVALUATION OF COLLABORATIVE FILTERING TECHNIQUES ON AMAZON FASHION DATA

Tóm tắt

Recommender systems play a crucial role in enhancing user experience by delivering personalized suggestions on e-commerce platforms. Among various techniques, collaborative filtering is widely adopted due to its ability to capture implicit relationships between users and items. However, the wide variety of available algorithms makes model selection challenging. This study conducts a comprehensive empirical evaluation of 18 collaborative filtering models categorized into three major groups: matrix factorization, heuristic/statistical methods, and deep learning, using the Amazon Fashion 2023 dataset. These models are assessed using AUC, NDCG@20, Precision@20, Recall@20, and training time. The results reveal that GMF achieves the highest AUC (0.7883), while VAECF and EASER demonstrate superior top-k recommendation quality, with Recall@20 reaching 0.1429. Conversely, models like BiVAECF and COE suffer from high training costs and limited effectiveness, especially in sparse data settings. The study highlights the strengths and weaknesses of each model group and recommends developing hybrid models that combine performance and efficiency for scalable recommender system deployment.