Feature selection for user-adaptive content-based music retrieval using Particle Swarm Optimization

Studies on content-based music retrieval (CBMR) which search music by analyzing their acoustic features and defining their similarity, have been conducted actively. However, it is desirable that the similarity evaluation be adaptive to each user's demand, because the search criteria differs user by user. In this paper, we propose a framework of CBMR that tries to satisfy the various demands of different users. We propose a method which improves retrieval accuracy to meet the demands of the users by adjusting the weights corresponding to the importance of features extracted from music using Particle Swarm Optimization (PSO). Moreover, we propose the use a type of PSO which enables an efficient parameter search by limiting the search domain according to the inherent characteristics of the parameter space of this problem. In the experiments, we verified that the suitable weight set is selected for different demands, improving the retrieval precision.