MUS490: Advanced Feature Selection and Unsupervised Clustering for Music Categorization

This paper presents MUS490, an innovative tool for the automated clustering and categorization of music files using unsupervised machine learning techniques. MUS490 leverages Mel-Frequency Cepstral Coefficients (MFCC), Onset-Based Spectral Information (OBSI), and Amplitude Modulation (AM) features to analyze timbre, genre, and instrumental characteristics of music datasets. The tool's primary contributions include the implementation of auto-k and auto-weight algorithms within the K-means framework, enabling dynamic determination of cluster counts and feature significance. Through a series of tests on diverse datasets, MUS490 demonstrates its capability to partition music into intuitive and sometimes unexpected categories, such as instrument type, genre, and tempo. While successful in many respects, the system's performance highlights challenges related to outlier detection and cluster precision, particularly in heterogeneous datasets. Future work aims to address these issues through improved heuristics, parallel processing, and integration of alternative clustering algorithms like DBSCAN. MUS490 represents a significant step towards fully automated music classification, with potential applications in music analysis, recommendation systems, and genre discovery.