Search

Increasing number of books published in a year and decreasing budgets have made collection development increasingly difficult in libraries. Despite the data to help decision making being available in the library systems, the librarians have little means to utilize the data. In addition, modern key technologies, such as machine learning, that generate more value out data have not yet been utilized in the field of libraries to their full extent. This study was set to discover a way to build a recommendation system that could help librarians who are struggling with book selection process. This thesis proposed a novel hybrid recommendation system for library book selection. The data used to build the system consisted of book metadata and book circulation data of books located in Joensuu City Library’s adult fiction collection. The proposed system was based on both rule-based components and a machine learning model. The user interface for the system was build using web technologies so that the system could be used via using web browser. The proposed recommendation system was evaluated using two different methods: automated tests and focus group methodology. The system achieved an accuracy of 79.79% and F1 score of 0.86 in automated tests. Uncertainty rate of the system was 27.87%. With these results in automated tests, the proposed system outperformed baseline machine learning models. The main suggestions that were gathered from focus group evaluation were that while the proposed system was found interesting, librarians thought it would need more features and configurability in order to be usable in real world scenarios. Results indicate that making good quality recommendations using book metadata is challenging because the data is high dimensional categorical data by its nature. Main implications of the results are that recommendation systems in domain of library collection development should focus on data pre-processing and feature engineering. Further invest