ENHANCING USER-BASED COLLABORATIVE FILTERING BY SIMILARITY COMPUTATION INCORPORATING POPULARITY TENDENCIES

Abstract

This study introduces a hybrid similarity measure for user-based collaborative filtering that combines traditional rating-based similarities with popularity-aware components to enhance neighborhood selection and prediction accuracy. Items are categorized into popular, diverse, and niche groups using a Pareto-based distribution of user ratings. Probabilistic user profiles are created to capture tendencies toward these categories, and similarities are computed using JensenShannon divergence. These category-based similarities are integrated with Pearson correlation through an adjustable α parameter, addressing sparsity challenges while preserving the precision of rating-based profiles. Experiments on three real-world datasets show that optimal performance is achieved at α=0.9, where rating-based similarities act as the primary driver of accurate predictions, while category-based profiles serve as supportive elements to refine neighborhood selection. The hybrid measure demonstrates significant improvements in MAE and RMSE, particularly in the sparsest dataset, where MAE is significantly reduced by 13.39% and RMSE by 17.35% compared to the baseline (α=1). This work highlights the hybrid measure’s ability to address sparsity while improving prediction accuracy. The inclusion of similarities based on user tendencies toward popular items further enhances neighborhood selection, contributing to more accurate and personalized recommendations across diverse data distributions.