Rahmawati Febrifyaning Tias, Triyanna Widiyaningtyas, Wahyu Sakti Gunawan Irianto, Wahyu Caesarendra
The performance of Collaborative Filtering (CF), which is commonly used in recommendation systems, often deteriorates under data sparsity and in the presence of cold-start users. To address this issue, this study proposes Wasserstein-User Profile Correlation Similarity (WUsim), a hybrid similarity model that combines Wasserstein Distance to capture similarity in rating distributions, with User Profile Correlation (UPC) to model behavioral proximity and user characteristics. This integration enables accurate similarity calculations even when co-rated items are limited. Evaluation on MovieLens-100K and MovieLens-1M using a random split (80:20) and a cold-start protocol demonstrates consistent improvements in rating prediction accuracy, measured by Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE). On MovieLens-100K, WUsim achieves a best RMSE of 1.083, while on MovieLens-1M the best RMSE is 1.025, and paired statistical significance testing (α = 0.05) confirmed that the observed improvements are statistically significant. Overall, these results indicate that the proposed hybrid similarity approach improves the robustness of CF against sparsity and cold-start, and generates more stable, informative, and efficient recommendations across various data scales. Copyright © by the authors
Department of Electrical Engineering and Informatics, Universitas Negeri Malang, Indonesia; Informatics Engineering, Faculty of Engineering, Universitas Bhayangkara Surabaya, Indonesia; Department of Mechanical and Mechatronics Engineering, Faculty of Engineering and Science, Curtin University Malaysia, Sarawak, Malaysia
