A Web-Based Book Recommendation System Implementing Machine Learning

Abstract

In this project, a personalized, precise book recommendation system is created based on the Book-Crossing dataset by using Truncated Singular Value Decomposition (SVD) of the matrices as a factorization on the dataset. The system addresses the problem of information overload in digital libraries, where the user has difficulty locating pertinent books within large collections of books. The preprocessing approach results in the development of a sparse user-book rating matrix, which is then SVD to reveal latent user-book relationships. Fuzzy matching, which is provided through the fuzzywuzzy library, makes it resilient to inaccurate input, whereas a web interface built with Streamlit presents recommendations with book cover images, making it more engaging to the user. Relevant suggestions are verified by qualitative appraisals and quick reactions to them (less than 3 seconds). The SVD is reported to have a 0.86822 RMSE on the Goodreads dataset [1], however quantitative measures (e.g. RMSE, MAE, F1-Score) of Book-Crossing are to be addressed in future work due to time constraints. SVD is more accurate and efficient in comparison with other methods such as ALS (RMSE: 1.09320, Goodreads dataset) [1] and user-based collaborative filtering [6]. The system encourages literacy, and is constrained by the cold start problem and collaborative filtering. The future development of the search will be on hybrid-filtering, real-time personalization, and cloud deployment that will allow the search to be better scaled and personalized and help facilitate educational and cultural development with improved book search.