ArXivDigest: A Living Lab for Personalized Scientific Literature Recommendation

Abstract

The purpose of this thesis is to explore different methods for recommending scientific literature to scientists and to explore different methods for doing topic extraction. We will update and use the already existing arXivDigest platform, which uses feedback from real users to evaluate article recommendations, to evaluate and compare these methods.

We introduce scientific literature recommendation methods based on term-based scoring, query expansion, semantic similarity and similar authors. While on topic recommendation we explore the RAKE and TextRank algorithms for topic extraction and TF-IDF weighting for topic similarity matching. These methods are all running live on the arXivDigest platform where we collect user feedback on the recommendations they provide.

We were able to get some users to sign up and use our platform, but they were unfortunately not active enough to generate sufficient interaction data by the time of submission to draw any reliable conclusions about system performance. We can however see that the arXivDigest platform is performing as it should and recommendations are submitted daily.