Music Recommendation from Song Sets

music analysis, information retrieval, multimedia indexing We motivate the problem of music recommendation based solely on acoustics from groups of related songs or 'song sets'. We propose four solutions which can be used with any acoustic-based similarity measure. The first builds a model for each song set and recommends new songs according to their distance from this model. The next three approaches recommend songs according to the average, median and minimum distance to songs in the song set. For a similarity measure based on K means models of MFCC features, experiments on a database of 18647 songs indicated that the minimum distance technique is the most effective, returning a valid recommendation as one of the top 5 32.5% of the time. The approach based on the median distance was the next best, returning a valid recommendation as one of the top 5 29.5% of the time.