Wiwik Suharso, Triyanna Widiyaningtyas, Ilham Ari Elbaith Zaeni, Muhammad Fatkhur Rizal, Salahudin Robo, Lukman Hakim
The Memory-Based Collaborative Filtering (MBCF) method uses implicit queries derived from user preferences. The method allows users to quickly and accurately obtain personalized recommendations. For this purpose, MBCF develops user-based and item-based techniques. Both techniques use previous rating data to measure the similarity between active users/items and target users/items. However, users exhibit different behavioral patterns when rating an item. Some users give high ratings for various items, while others give low ratings for the same items. These differences reflect the user's rating behavior. We compared three similarity measures: Cosine, Adjusted Cosine (ACosine), and Pearson. The measures were used in mean-centered prediction methods under user-based and item-based, respectively. In practice, normalization of mean-user ratings and mean-item ratings is applied to the ACosine and Pearson calculation formulas, respectively. All experiments used the original MovieLens 100K dataset. The experimental results show that Cosine, Pearson, and ACosine under item-based obtained the lowest error, respectively. The evaluation metrics show RMSE values of 0.988, 1.002, and 1.005, respectively. Meanwhile, the MAE values are 0.773, 0.785, and 0.785, respectively. Item-based techniques outperform user-based techniques in three similarity measures. © 2025 IEEE.
Universitas Negeri Malang, Electrical Engineering and Informatics, Malang, Indonesia; Universitas Hasyim Asy'ari, Information Technology, Jombang, Indonesia; Universitas Yapis Papua, Information Systems, Papua, Indonesia; Politeknik Negeri Jember, Digital Business, Jember, Indonesia
