DESIGNING A WELL-BALANCED TWO-DIMENSIONAL RECOMMENDER SYSTEM FOR PUBLIC LIBRARIES

Abstract

This thesis investigates the steps and decisions needed to design a well-balanced, twodimensional recommender system for public libraries. Where they exist, recommender systems for public libraries are one-dimensional in nature, and not all products offer personalized recommendations. Large commercial platforms like Netflix, that provide digital media content, have been using two-dimensional recommendations for a long time. In contrast, a two-dimensional option to navigate the public library collection was not found when exploring the current state of recommender solutions for public libraries.
The research has three key areas: 1) identification of recommender algorithms and public library data features that provide accurate recommendations, 2) strategies for selecting carousels (first dimension) and carousel-items (second dimension) for a user, and 3) improvements on the beyond-accuracy aspects diversity, serendipity and novelty within the item selection, to make the system well-balanced.
The research is scoped to public libraries in theUnited States of America that use OCLC’s Wise Library Management System. The data that is processed contains implicit feedback in the form of user borrowing history, and bibliographic metadata. Other user-item interactions, such as item ratings, item reviews, or click-through logs, are excluded from this research. Algorithms froma publicly available software repositorywere included, and without modification: no tweaking was done, and no new recommender algorithms were created.
Offline experiments were conducted using a selection of available algorithms and library data to obtain item prediction scores for users. Using the design science methodology, prototype software was implemented to select carousels and carousel-items, after which its behavior was analyzed. Additional offline experiments were conducted to measure the impact on beyond-accuracy aspects when using different item selection strategies. Using the collaborative filtering algorithm LightGCN, a recommender model was trained that generates item predictions for users that, through a qualitative review, are believed to match user interests. Considering a user’s top-predicted items and most-recent transaction items, common item features in both item sets can be determined for that user, and then applied to select carousels. Using a combined strategy that distributes its focus over diversity, serendipity, and novelty, I found that this results in a well-balanced selection of items for the selected carousels.

Date of Award	19 Dec 2024
Original language	English
Supervisor	Gideon Maillette de Buij Wenniger (Examiner), Arjen Hommersom (Co-assessor), Emil Poortman (External assessor) & Paul Lucassen (External assessor)