Cliques in Music Recommendation Engines: An Experiment Using Last.fm and Spotify Artist Similarity Networks

Abstract

With the influx of free music available to the public, there is an ever-growing need for effective music recommendation engines (also known as recommender systems). When evaluating items { music as well as other entities { users often have an over- whelming number of options, and most users lack the expertise to discern which item would be best to use or purchase. Recommender systems provide suggestions to users to aid in the decision-making process of selecting items. While usage of such systems is expanding, there are many weaknesses in their methodology, chiefly in the over-emphasis of prediction accuracy as the key indicator of success. In order to im- prove recommendation quality and guard against potential e ects of popularity bias, other factors, such as diversity of recommendations, would be invaluable to evaluate. Here, we assess two music recommendation engines, Last.fm and Spotify. This paper outlines recommendation problems, obstacles unique to music recommender systems, conventional approaches used by music recommendation engines, and some strengths and weaknesses of various methods. An experiment with artist similarity data from the Last.fm and Spotify music recommendation platforms is presented. Two artist similarity networks were made using the engines open-source APIs. A comparative analysis of these networks using the global clustering coe cient leads us to conclude that popularity bias is less prevalent in systems with types of listeners who have greater interest in music.