Automatic Outlier Detection in Music Genre Datasets

Chang-Tien Lu

Abstract

Outlier detection, also known as anomaly detection, is an important topic that has been studied for decades. An outlier detection system is able to identify anomalies in a dataset and thus improve data integrity by removing the detected outliers. It has been successfully applied to different types of data in various fields such as cyber-security, finance, and transportation. In the field of Music Information Retrieval (MIR), however, the number of related studies is small. In this paper, we introduce different state-of-the-art
outlier detection techniques and evaluate their viability in the context of music datasets. More specifically, we present a comparative study of 6 outlier detection algorithms applied to a Music Genre Recognition (MGR) dataset. It is determined how well algorithms can identify mislabeled or corrupted files, and how much the quality of the dataset can be improved. Results indicate that state-of-the-art anomaly detection systems have problems identifying anomalies in MGR datasets reliably.