Evaluation of modeling music similarity perception via feature subset selection

Abstract

In this paper, we describe and discuss the evaluation process and results of a content-based music retrieval system that we have developed. In our system, user models embody the ability of evolving and using different music similarity measures for different users. Specifically, a user-supplied relevance feedback and related neural network-based incremental learning procedures allows our system to determine which subset of a set of objective acoustic features approximates more efficiently the subjective music similarity perception of an individual user. The evaluation results verify our hypothesis of a direct relation between subjective music similarity perception and objective acoustic feature subsets. Moreover, it is shown that, after training, retrieved music pieces exhibit significantly improved perceived similarity to user-targeted music pieces. © Springer-Verlag Berlin Heidelberg 2007.