Book Recommendation System Using Similarity-Based Collaborative Filtering Approach

Abstract

This study implements and compares the performance of three primary approaches to book recommendation systems: rank-based methods, similarity-based collaborative filtering (both user-user and item-item), and matrix factorization-based collaborative filtering. The dataset comprises 433,671 user ratings from 78,805 users on 185,973 books, enriched with book metadata such as title, author, publication year, and publisher. The recommendation systems were developed using the Surprise library, with hyperparameter optimization performed via grid search cross-validation. Model performance was evaluated using precision@k, recall@k, and F1-score metrics, as well as RMSE for prediction accuracy. Results indicate that the user-user similarity-based collaborative filtering model achieved the best performance in terms of relevance, attaining an F1-score of 0.86. This model effectively identifies users with similar preferences and recommends books based on collective behavior patterns. Meanwhile, the matrix factorization approach yielded the lowest RMSE value of 1.50, highlighting its strength in capturing latent factors that influence user preferences. The item-item similarity model also showed reasonable performance but did not surpass the other approaches, possibly due to homogeneity in item rating patterns across users. Overall, the study confirms that user-user similarity is highly effective for datasets exhibiting consistent user behavior, while matrix factorization excels in minimizing prediction error by leveraging latent feature structures. These findings offer valuable insights for developing adaptive recommendation systems in book-centric literacy platforms and content-driven e-commerce applications.