Optimization of User-Based Collaborative Filtering Movie Recommendation System Using Mean-Centering and Overlap Weighting on Cosine Similarity

Abstract

Rapid technological advancements have changed the way people enjoy entertainment. Viewers can now watch movies online without having to go to the cinema. The rapid development of online movie streaming services has increased the need for systems capable of providing content recommendations based on user preferences. This study aims to optimize a User-Based Collaborative Filtering (UBCF)-based recommendation system using the MovieLens 100K dataset containing 100,000 ratings from 943 users for 1,682 movies. The evaluation was conducted by dividing the data into 80% training data and 20% testing data. The system was optimized through two main approaches: user average normalization (mean-centering) in the KNN method to reduce bias in rating scale differences between users, and the application of overlap weighting on cosine similarity to give greater weight to user pairs with a greater number of shared item ratings. Based on experiments, a value of k = 50 was chosen as the optimal trade-off point in predicting ratings. The experimental results show that KNN with mean centering (KNNWithMeans) consistently outperforms standard KNN. The cosine + KNNWithMeans model produced an RMSE of 0.9701 and an MAE of 0.7567, lower than cosine + KNN (RMSE 1.0377; MAE 0.8226). Further overlap weighting was shown to improve prediction accuracy, with the combination of weighted cosine with α = 1 and KNNWithMeans providing the best performance with an RMSE of 0.9686 and an MAE of 0.7556.