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.
