Thomas Rönnberg Classification of Heavy Metal Subgenres with Machine Learning Master’s Thesis in Information Systems Supervisors: Dr. Markku Heikkilä Asst. Prof. József Mezei Faculty of Social Sciences, Business and Economics Åbo Akademi University Åbo 2020 Subject: Information Systems Writer: Thomas Rönnberg Title: Classification of Heavy Metal Subgenres with Machine Learning Supervisor: Dr. Markku Heikkilä Supervisor: Asst. Prof. József Mezei Abstract: The music industry is undergoing an extensive transformation as a result of growth in streaming data and various AI technologies, which allow for more sophisticated marketing and sales methods. Since consumption patterns vary by different factors such as genre and engagement, each customer segment needs to be targeted uniquely for maximal efficiency. A challenge in this genre-based segmentation method lies in today’s large music databases and their maintenance, which have shown to require exhausting and time-consuming work. This has led to automatic music genre classification (AMGC) becoming the most common area of research within the growing field of music information retrieval (MIR). A large portion of previous research has been shown to suffer from serious integrity issues. The purpose of this study is to re-evaluate the current state of applying machine learning for the task of AMGC. Low-level features are derived from audio signals to form a custom-made data set consisting of five subgenres of heavy metal music. Different parameter sets and learning algorithms are weighted against each other to derive insights into the success factors. The results suggest that admirable results can be achieved even with fuzzy subgenres, but that a larger number of high-quality features are needed to further extend the accuracy for reliable real-life applications. Keywords: Artificial Intelligence, Machine Learning, Deep Learning, Supervised Learning, Predictive Analytics, Data Science, Data Mining, Music Information Retrieval, Automatic Music Genre Classification, Digital Signal Processing, Audio Signal Processing, Computational Musicology, Spectrograms, Heavy Metal Date: Number of pages: 120 Table of Contents List of Acronyms ................................................................................................................................. 1 1. Introduction ...................................................................................................................................... 3 1.1 Background ................................................................................................................................ 3 1.2 Problematization ........................................................................................................................ 5 1.3 Objective .................................................................................................................................... 7 1.4 Method ....................................................................................................................................... 7 1.5 Previous research ....................................................................................................................... 9 1.6 Disposition ................................................................................................................................. 9 2. Feature Extraction and Music Information Retrieval .................................................................... 11 2.1 Music Representation Methods ................................................................................................ 11 2.2 Audio Signals and Waveforms ................................................................................................ 11 2.3 Short-Time Fourier Transform and Spectrograms ................................................................... 14 2.4 Low-Level Feature Extraction ................................................................................................. 17 2.4.1 Mel-frequency Cepstral Coefficients ................................................................................ 18 2.4.2 Spectral Centroid............................................................................................................... 19 2.4.3 Spectral Bandwidth ........................................................................................................... 19 2.4.4 Spectral Roll-off ................................................................................................................ 19 2.4.5 Root-Mean-Square Energy ............................................................................................... 19 2.4.6 Zero-crossing Rate ............................................................................................................ 20 2.4.7 Dynamic Tempo ................................................................................................................ 20 2.5 Aggregation Methods ............................................................................................................... 20 2.6 Genre Taxonomy...................................................................................................................... 20 2.6.1 Heavy Metal ...................................................................................................................... 22 2.6.2 Thrash Metal ..................................................................................................................... 22 2.6.3 Death Metal ....................................................................................................................... 23 2.6.4 Black Metal ....................................................................................................................... 24 2.6.5 Folk Metal ......................................................................................................................... 25 3. Classification and Machine Learning ............................................................................................ 27 3.1 Learning types in Machine Learning ....................................................................................... 27 3.1.1 Supervised Learning ......................................................................................................... 27 3.1.2 Unsupervised Learning ..................................................................................................... 28 3.1.3 Reinforcement Learning ................................................................................................... 28 3.2 Model Building ........................................................................................................................ 29 3.2.1 Feature Engineering .......................................................................................................... 29 3.2.2 Data Exploration and Preprocessing ................................................................................. 29 3.2.3 Model Selection and Parameter Tuning ............................................................................ 31 3.2.4 Model Validation and Resampling Methods .................................................................... 31 3.2.4.1 Validation Set Approach ............................................................................................ 32 3.2.4.2 K-Fold Cross-Validation ............................................................................................ 32 3.2.4.3 Bootstrapping ............................................................................................................. 34 3.2.5 The Bias-Variance Tradeoff .............................................................................................. 34 3.2.6 The Curse of Dimensionality ............................................................................................ 35 3.2.6.1 Feature Selection ........................................................................................................ 35 3.2.6.2 Principal Component Analysis................................................................................... 36 3.3 Performance Metrics for Model Evaluation ............................................................................. 36 3.3.1 Classification Accuracy .................................................................................................... 36 3.3.2 Confusion Matrix .............................................................................................................. 37 3.3.3 Precision, Recall, F1 .......................................................................................................... 37 3.4 Learning algorithms ................................................................................................................. 38 3.4.1 Naïve Bayes ...................................................................................................................... 39 3.4.2 K-Nearest Neighbors ......................................................................................................... 40 3.4.3 Decision Trees................................................................................................................... 41 3.4.4 Support Vector Machines .................................................................................................. 43 3.4.5 Random Forests................................................................................................................. 44 3.4.6 AdaBoost ..........................................................................................................................