Aggregating Music Recommendation Web APIs by Artist Brandeis Marshall Computer and Information Technology Purdue University brandeis@purdue.edu Abstract Through user accounts, music recommendations are refined by user-supplied genres and artists pref- erences. Music recommendation is further compli- cated by multiple genre artists, artist collaborations and artist similarity identification. We focus pri- marily on artist similarity in which we propose a rank fusion solution. We aggregate the most simi- lar artist ranking from Idiomag, Last.fm and Echo Nest. Through an experimental evaluation of 300 artist queries, we compare five rank fusion algo- rithms and how each fusion method could impact the retrieval of established, new or cross-genre mu- sic artists. Keywords: rank aggregation methods, artist sim- ilarity, music information retrieval 1. Introduction Online radio can be accessed in one of two ap- proaches: subscription such as Sirius Satellite Ra- dio and XM Satellite Radio and free such as AOL Radio, Pandora, Last.fm, Idiomag and Echo Nest. Each online radio portal allows music listeners to create a user account in hopes of tracking music genre and artist preferences. In most cases, the user chooses a radio station with a programmed playlist. In contrast, Pandora only needs a single music artist to being the customized user playlist. If the user would like to listen to different music genres, she must provide a sample music artist for Pandora to generate an appropriate playlist. Some challenges facing music recommendation are multiple genre artists, music artist collabora- tions and artist similarity identification. Many mu- sic artists can be classified in more than one genre due to the artists deciding to alter her sound or assessing the influences of one genre onto another genre. Music artist collaborations have become pop- ular in which collaborations are within and across music genre. These collaborations may occur on more than one song or album. For two artist col- laborations, a music listener may like the song col- laboration but only enjoy the music from one of the artists. Artist similarity is primarily user-driven since likeness is highly subjective. These challenges, therefore, make capturing artist similarity difficult. In this paper, we are concerned with identifying similar artists, which serves as a precursor to how music recommendation can handle the more com- plex issues of multiple genre artists and artist col- laborations. We consider the individual most sim- ilar artist ranking from three public-use Web APIs (Idiomag, Last.fm and Echo Nest) as different per- spectives of artist similarity. We examine the level of overlap amongst these Web APIs through rank fusion methods. By understanding this overlap, we can more easily isolate the multiple genre artists and artist collaborations. The specific contributions of this paper are: (1) examine rank fusion as a solution to artist similarity and (2) perform a quantitative study of artist sim- ilarity using five fusions algorithms including Av- erage, Condorcet-fuse, CombMNZ, PageRank and Median. 2. Related Work Recommendation systems have been developed for niche domains such as books, movies and mu- sic [4, 9, 13]. Collaborative filtering has become the accepted approaches in order to provide user- specific results using information from many users. However, the prior work of [4, 13] concentrate on the users’ playlist through song properties includ- ing pitch, duration and loudness. The music genre, on the other hand, is more complex than the users’ playlist. We focus on music genre because of its song and artist diversity while avoiding the pre- processing of song properties conducted in prior re- search. In recommending music using text, labels or tags