Learning Song Similarity from Radio Station Playlists


This extended abstract details a submission to the 2009 Music Information Retrieval eXchange (MIREX) audio music similarity and retrieval task. We propose a method for computing song similarity by learning song transition probabilities from audio features extracted from songs played in professional radio station playlists. We train a binary classifier to distinguish between 2-song sequences that were played on the radio and sequences that were not. The certainty of the trained model when presented with a song sequence can then be interpreted as the similarity between the two songs. The model used is explained in more details in [1]. 1. LEARNING FROM RADIO STATION PLAYLISTS 1.1 Defining a playlist Our similarity model is trained on professional radio station playlists. For this experiment, we consider a playlist to be a sequence of 2 consecutive plays uninterrupted by a commercial break. Suppose a radio station plays the tracks ta, tb and tc one after the other, we will consider {ta, tb} and {tb, tc} as two 2-song sequences ∈ S. We consider the sequences {ta, tb} and {tb, ta} as two distinct sequences. The model’s output will thus be non-symmetric in regards to the order in which the songs are presented. 1.2 Playlist sources We used playlist data from two sources. First, we used data from the free Internet-streamed radio station RadioParadise 1 , who provided us with 575 days worth of data. The data consists of 195,692 plays, 6,328 unique songs and 1,972 unique artists. We also used Yes.com’s API 2 , which gives access to the play history of thousands of radio stations in the United 1 http://www.radioparadise.com 2 http://api.yes.com Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. c © 2009 International Society for Music Information Retrieval. States. We mined data for 57 days, totaling in 449 stations, 6,706,830 plays, 42,027 unique songs and 9,990 unique artists. 1.3 Putting the data together Combining all the data yielded 6,902,522 plays, with an average of 15,338 plays per station. Since our model uses audio features as input, we need the audio file for all the songs we will use. Of the 47,044 total songs played in the playlists we used, we were able to obtain the audio files for 7,127 tracks. This reduced the number of distinct usable 2-song sequences to 180,232. The sequences for which we had all the audio files were combinations from 5,562 tracks. We did not possess a set of explicit negative examples (i.e. two-song sequences that a radio station would never play). In order to train the model with examples from both the positive and negative class, we considered any song sequence that was not observed as a negative example. During training, at each new epoch, we randomly sampled a new set of negative examples matched in size to our positive example set. 2. SIMILARITY MODEL OVERVIEW

Cite this paper

@inproceedings{Maillet2009LearningSS, title={Learning Song Similarity from Radio Station Playlists}, author={François Maillet and Douglas Eck}, year={2009} }