Comparing User Rating-Based Similarity to User Behavior-Based Similarity in Movie Recommendation Systems

Abstract

Implementing similarity functions will affect how well the recommender system performs. The traditional similarity solely considers user rating values to calculate the similarity between users. Some studies have recently developed improved similarity functions involving user behavior values. These algorithms combine similarity depending on the user ratings with similarity considering the user behavior. This study aims to compare the performance of the similarity depending on the user rating and the similarity considering the user behavior in movie recommender systems. The similarity functions applied are Cosine similarity, User score Probability Collaborative Filtering (UPCF), User Profile Correlation-based Similarity (UPCSim), and Clustering-Based UPCSim (CB-UPCSim). The experiment results in MovieLens 100K and MovieLens 1M show that the similarities considering the user behavior (UPCF, UPCSim, and CB-UPCSim) can reduce the Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) values compared to the similarity only depending on the user rating (Cosine). In addition, the CB-UPCSim showed the fastest running time and the highest F1 score, recall, and precision compared to the UPCF, UPCSim, and Cosine similarity. It indicates that the CB-UPCSim approach performs better than the three other similarity functions.