Content-Based Music Retrieval with Nonlinear Feature Space Transformation using Relevance Feedback

In recent years, studies of similar music retrieval have been conducted actively. However, because the similarity of music is based on subjective measures, the systems need to be adaptive to user preference. In this paper, we propose an effective method for adaptive similar music retrieval reflecting the user preference by nonlinear feature space transformation based on relevance feedback. The user's evaluation to initial retrieval ranking is used to train a neural network feature space transformation. Also, as the initial stage of retrieval, a coarse division is made to the set of music in the database, to reduce the retrieval computation. In the experiments, it was observed that in the retrieval after feature space transformation, the proposed method gave preferred retrievals according to objective measures when compared with the case without transformation.