Portal de Periódicos da CAPES

Reliable Book Recommender System: An Evaluation and Comparison of Collaborative Filtering Algorithms

2021; Springer International Publishing; Linguagem: Inglês

10.1007/978-3-030-90055-7_20

2367-3370

Aldin Kovačević, Zerina Mašetić,

Topic Modeling

Nowadays, a vast number of online businesses and services use recommender systems to provide relevant product suggestions to their customers. Recommender systems can be defined as systems employing various data mining techniques to predict users’ interest in certain items or topics, among a tremendous amount of available data. While there exist varying methodologies to construct a recommender system, one of the most potent and commonly used techniques is collaborative filtering (CF), which evaluates items based on the opinions of other users. This paper examines several commonly used collaborative filtering models with the goal of creating a reliable book recommendation system. The research focuses on training, evaluating and comparing several neighborhood-based and matrix factorization-based recommender algorithms, using a combination of two book datasets: the BookCrossing dataset compiled in 2004, and the Goodbooks “most popular books” dataset from 2017. Grid search with cross-validation is used to find the best performing combination of several hyper-parameters and different CF models. Ultimately, a matrix factorization-based SVD++ approach is selected as the most performant model, with MAE of 0.57 and RMSE of 0.752, making it a considerably good model based on the provided data. Moreover, a manual comparison between a user’s book history and their recommended book titles shows that users are generally recommended books of the same or similar genres. Lastly, the paper discusses certain shortcomings of the proposed model and offers several suggestions for future improvements.