On mixture model complexity estimation for music recommender systems

Content-based music navigation systems are in need of robust music similarity measures. Current similarity measures model each song with the same model parameters. We propose methods to efficiently estimate the required number of model parameters of each individual song. First results of a study on relationships between a small set of basic audio features are presented. We conclude that there are only very small correlations between models on low- and on high-dimensional features. When we compare a very simple clustering algorithm with an algorithm that estimates model parameters using the MDL criterium, we find a surprisingly strong correlation between the estimated number of mixture components.