Effective Nearest-Neighbor Music Recommendations

Abstract

Automated recommendations for next tracks to listen to or to include in a playlist are a common feature on modern music platforms. Correspondingly, a variety of algorithmic approaches for determining tracks to recommend have been proposed in academic research. The most sophisticated among them are often based on conceptually complex learning techniques which can also require substantial computational resources or special-purpose hardware like GPUs. Recent research, however, showed that conceptually more simple techniques, e.g., based on nearest-neighbor schemes, can represent a viable alternative to such techniques in practice. In this paper, we describe a hybrid technique for next-track recommendation, which was evaluated in the context of the ACM RecSys 2018 Challenge. A combination of nearest-neighbor techniques, a standard matrix factorization algorithm, and a small set of heuristics led our team KAENEN to the 3rd place in the "creative" track and the 7th one in the "main" track, with accuracy results only a few percent below the winning teams. Given that offline prediction accuracy is only one of several possible quality factors in music recommendation, practitioners have to validate if slight accuracy improvements truly justify the use of highly complex algorithms in real-world applications.