Music classification by transductive learning using bipartite heterogeneous networks

Abstract

The popularization of music distribution in electronic format has increased the amount of music with incomplete metadata. The incompleteness of data can hamper some important tasks, such as music and artist recommendation. In this scenario, transductive classification can be used to classify the whole dataset considering just few labeled instances. Usually transductive classification is performed through label propagation, in which data are represented as networks and the examples propagate their labels through their connections. Similarity-based networks are usually applied to model data as network. However, this kind of representation requires the definition of parameters, which significantly affect the classification accuracy, and presentes a high cost due to the computation of similarities among all dataset instances. In contrast, bipartite heterogeneous networks have appeared as an alternative to similarity-based networks in text mining applications. In these networks, the words are connected to the documents which they occur. Thus, there is no parameter or additional costs to generate such networks. In this paper, we propose the use of the bipartite network representation to perform transductive classification of music, using a bag-of-frames approach to describe music signals. We demonstrate that the proposed approach outperforms other music classification approaches when few labeled instances are available.