Music Recommender Performance for Beyond-Mainstream Listeners

Now that digital content consumption and therefore also music streaming is more prominent than ever, music recommender systems have become essential for improving user experience. While traditional recommender systems perform well for mainstream listeners, they often fall short for beyond mainstream listeners. Using the Last.fm - Beyond Mainstream (LFM-BeyMS) dataset, which includes listening histories and track metadata, this thesis analyzes various different state-of-the-art music recommender systems and evaluates them on different performance metrics, particularly for the less predictable tastes of beyond mainstream listeners. When comparing collaborative filtering, content-based filtering, and hybrid filtering approaches, we found that a deep learning-based hybrid recommender system offered the best recommendation quality for beyond  mainstream listeners, even outperforming the mainstream listener group.