Abstract

With the advent of online music platforms, users experience an ever increasing volume of available music offers. As a consequence, users may feel overwhelmed. Recommender systems direct the attention of users to a small subset of music tracks that most likely match the taste of the users. This thesis focuses on sequential music recommendation and in particular on session-completion, i.e., the task of recommending the remainder of a session given the knowledge of its initial segment. Music listening sessions often include repeated tracks. Research shows that modeling such individual relistening behavior with models of the human memory may be promising. In particular, Adaptive Control of Thought Rational (ACT-R) outperforms some strong baselines in session-completion. However, since recommender systems relying on ACT-R only can not make use of information encoded in the listening history of other users, these are unable to recommend novel tracks to users. In contrast, the main approach to recommender systems, collaborative filtering (CF), bases recommendations on listening histories of other similar users rather than on individual relistening behavior. CF-based recommender systems may recommend novel items, but their recommendations often lack explainability. A combination of CF and ACT-R may lead to novel and explainable recommendations combining individual relistening and listening behavior of other users. However, no prior research combining CF and ACT-R exists. To close this research gap, this thesis proposes several hybrid algorithms that integrate CF with ACT-R. We provide three major contributions to the current research in the recommender system domain. First, we develop several new algorithms that combine ACT-R with CF for the task of session-completion. Second, we provide an extensive analysis of the performance of these algorithms on the large LFM-2b dataset of Last.fm listening histories, in comparison with well-established baselines and state-of-the-art models for the task of music session-completion. Performance is measured in terms of accuracy and beyond-accuracy metrics such as novelty, diversity, and popularity miscalibration of recommendations. In particular, we empirically identify for each beyond-accuracy metric an algorithm that performs best for it. Third, we demonstrate how some of the proposed algorithms can be used to explain music recommendations. Since each component of the memory module of ACT-R is designed to model a different aspect of human memory, the recommendations provided by the proposed algorithms are explainable. Additionally, we show how the explanations can be beneficial for users, platform providers, and producers.

Citation

BibTeX

@misc{Wallmann2023master-thesis,
    title = {Explainable Sequential Music Recommender Systems: Integrating Collaborative Filtering and Theories of Human Memory},
    author = {Wallmann, Christian},
    school = {Johannes Kepler University Linz},
    year = {2023}
}