A Novel Genre-Specific Feature Reduction Technique through Association Analysis
Adam Lefaivre, Yingsheng Zhang · DOAJ (DOAJ: Directory of Open Access Journals) · 2020
We consider the genre classification problem in Music Information Retrieval and report our initial investigation on reducing the number of features that are used in genre classification. Each music genre has its own characteristics, which distinguish it from other genres. We adapt association analysis to capture those characteristics using acoustic features, i.e., each genre's characteristics are represented by a set of features and their corresponding values. Our goal is to select the ""most representative"" features for each genre. Such features are unique in distinguishing a genre and therefore should be singled out. We propose two criteria for comparing and selecting those unique features of each genre. The details of our proposed approach are presented. The effectiveness of our approach is demonstrated and discussed through empirical experiments.