Exploring musical relations using association rule networks

Abstract

Music information retrieval (MIR) has been gaining increasing attention in both industry and academia. While many algorithms for MIR rely on assessing feature subsequences, the user normally has no resources to interpret the significance of these patterns. Interpreting the relations between these temporal patterns and some aspects of the assessed songs can help understanding not only some algorithms’ outcomes but the kind of patterns which better defines a set of similarly labeled recordings. In this work, we present a novel method to assess these relations, constructing an association rule network from temporal patterns obtained by a simple quantization process. With an empirical evaluation, we illustrate how we can use our method to explore these relations in a varied set of data and labels.