Collaborative filtering models an experimental and detailed comparative study.

Area of Science:

• Computer Science
• Artificial Intelligence
• Data Science

Background:

• Recommender systems (RS) personalize user experiences across domains like e-commerce and entertainment.
• Collaborative Filtering (CF) is a key RS technique, utilizing user similarity to recommend items.
• Existing CF methods include memory-based, model-based, and neural network approaches.

Purpose of the Study:

• To conduct an experimental comparative analysis of various collaborative filtering recommender system methods.
• To evaluate the performance of different CF techniques on benchmark datasets using multiple metrics.
• To provide insights into the strengths, limitations, and practical applicability of each method.

Main Methods:

• Comparative analysis of memory-based (KNN), model-based (SVD, SVD++, co-clustering), and neural network (NCF, DeepFM, LightGCN) CF methods.
• Evaluation on MovieLens datasets (100K, 1M, 25M) using metrics like RMSE, MAE, NDCG@10, and Precision@10.
• Detailed examination of the working mechanisms, advantages, and disadvantages of each model.

Main Results:

• Neural and graph-based models show significant improvements (up to 15% ranking gains) on large datasets for rating accuracy and top-k ranking.
• Simpler methods (KNN, SVD) remain effective for smaller datasets or low-resource scenarios due to ease of implementation and interpretability.
• Performance gains vary based on dataset size, model complexity, and evaluation metrics.

Conclusions:

• The choice of CF technique requires balancing computational cost, scalability, and model intricacy.
• Neural and graph-based methods offer superior performance on large-scale data, while traditional methods provide a practical baseline.
• Findings offer practical guidance for selecting appropriate recommender system techniques based on specific application needs and data characteristics.