Collaborative Filtering In Recommender Systems: A Comparative Evaluation

Abstract

Recommender systems have become a core component of digital platforms, enabling personalized content delivery and enhancing user engagement across industries such as e-commerce, entertainment, and social networking. This study presents a comparative evaluation of collaborative filtering (CF) algorithms, including User-Based CF, Item-Based CF, Singular Value Decomposition (SVD), Non-negative Matrix Factorization (NMF), and a Hybrid CF model that integrates both memory-based and model-based approaches. Using the MovieLens 1M and Amazon Product Review datasets, the algorithms were assessed across multiple performance dimensions; predictive accuracy (RMSE, MAE), ranking quality (Precision@10, Recall@10, F1-score), and computational efficiency (training time, scalability, and memory usage). The results revealed that the Hybrid CF model achieved the highest overall performance, with the lowest RMSE (0.811) and the highest F1-score (0.731), outperforming traditional CF methods by over 13% in accuracy and 11% in ranking metrics. Although it required slightly higher computational resources, the Hybrid CF model demonstrated superior scalability and adaptability in both sparse and dense data environments. Cluster analysis further confirmed a clear distinction between memory-based and model-based methods, with the Hybrid CF forming an independent sub-cluster due to its unique performance balance. The findings conclude that hybrid collaborative filtering represents the most effective and future-ready solution for modern recommender systems, offering an optimal blend of accuracy, efficiency, and scalability for large-scale, data-intensive applications.