Report.ipynb describes the end-to-end machine learning workflow including exploratory data visualizations, feature preprocessing, model selection, experiments and results, and analysis.

To tackle the problem of classifying the musical genre of songs based on "carefully chosen features", approaches using support vector machines (SVM), Random Forest, XGBoost, and deep neural networks were used.

Data analysis showed high class imbalance in the dataset, and visualizations such as the correlation matrix of features, 3-dimensional PCA projection, and feature-class plots, indicate that many of the features are possibly non-informative and that classes are not easily separable.

On the Kaggle dataset, the best variants of all models ranged from 60.2%-66.2% in accuracy (33rd out of 406) and 0.236-0.171 in log-loss (38th out of 371), with XGBoost scoring the highest on both metrics.

Required Python packages are seaborn, numpy, scikit-learn, matplotlib, pandas

Running python classifiers.py will train and evaluate a given model, construct Kaggle submissions, and save various plots (e.g. confusion matrix, learning curve) and pickled models to files.

Args:

• --model, -m: Can be XGBoost (default), RandomForest, SVC, or KNN.
• --scale: Scale data with min-max scaling.
• --drop: Drop non-informative MFCC features (216-219).
• --final: Use all training data when training the model (default=False)
• --remarks, -r: [optional] A string to be saved in the filenames of exported resources.
• --neighbors. If the model is KNN, an integer defining the number of neighbours.

An example usage is python classifiers.py --model "SVC" --scale --drop --remarks "early stopping"

Voting classifiers ensembling several models can be run with python ensemble.py, with arguments:

• --model, -m: ensemble_rf_xgb (default), ensemble_rf_xgb_svc, final_ensemble_rf_xgb, final_ensemble_rf_xgb_svc, ensemble_rf_xgb_unfitted, ensemble_rf_xgb_svc_unfitted.
• --final: Use all training data when training the model.

python gridsearch.py