An Unsupervised Approach to Anomaly Detection in MusicrnDatasets

摘要

This paper presents an unsupervised method for systematicallyrnidentifying anomalies in music datasets. The modelrnintegrates categorical regression and robust estimation techniquesrnto infer anomalous scores in music clips. When appliedrnto a music genre recognition dataset, the new method is ablernto detect corrupted, distorted, or mislabeled audio samplesrnbased on commonly used features in music information retrieval.rnThe evaluation results show that the algorithmrnoutperforms other anomaly detection methods and is capablernof fnding problematic samples identifed by human experts.rnThe proposed method introduces a preliminary frameworkrnfor anomaly detection in music data that can serve as a usefulrntool to improve data integrity in the future.