This repository studies the performance of various regressor models on the prediction of progression of the Parkinson’s disease’s total and motor UPDRS metrics. The studied models are Random Forest Regressor (RFR), Support Vector Regressor (SVR), Multi Layer Perceptron networks (MLP), Gradient Boosting (GBR) and Adaptive Neuro-fuzzy Inference System (ANFIS). These models where tested on a variety of configurations, involving the use of Principal Component Analysis (PCA) and Recursive Feature Elimination (RFE) for dimensionality reduction, and the use of clustering and ensembling techniques.
More information on the results and the experimental setup can be found in the attached project report.
This repository uses the Parkinson's Telemonitoring Dataset from the UCI repository. For training of the models all features from the dataset but subject# and test_time are used.
The scripts that can be executed in order to run our experiments can be found under the experiments folder:
- hyperparameters_search: search of the hyperparameters for each model.
- models_all_dataset: regression using different models and the original dataset.
- models_projected_dataset: regression using different ensembled models and the projected datasets with Clustering + PCA.
- models_recursive_feature_elimination: regression using GBR and RFR models, with recursive feature elimination (RFE).
- reduction: inside there is the script that projects (reduces) the dataset using PCA and the different clustering algorithms.
These scripts print an output with their results and save them in the folder results. Under utils you can find scripts that, with the results obtained by the experiments, generate plots comparing the different algorithms. The plots are saved in media. The project was developed with Python 3 (we recommend Python >= 3.6).