Comparison of Riemannian Machine Learning and Deep Learning Methods for P300 Signals Using the MOABB Framework

This repository contains the source code, experimental results, and reproduction notebooks for the study: "Comparison of Riemannian Machine Learning and Deep Learning Methods for P300 Signals Using the MOABB Framework". Overview This study compares seven classifiers for P300 EEG signal classification using the BNCI2014_009 dataset from the MOABB framework. The evaluated models include: 4 Riemannian Machine Learning pipelines: TGSP+SVM, TGSP+LDA, TGSP+LR, and RMDM. 3 Deep Learning architectures: EEGNet, ShallowConvNet, and DeepConvNet. Experiment Setup Dataset: BNCI2014_009 (10 subjects, 3 sessions, 16 channels, 256 Hz) Trial Scenarios: Very Low (5), Low (10), Medium (20), and Full (864 trials/class) Evaluation: Within-session (5-fold Cross-Validation) and Cross-session (Leave-One-Subject-Out / LOSO) Statistical Test: Wilcoxon signed-rank ($\alpha = 0.05$) Key Findings Performance: Deep Learning significantly outperforms Riemannian ML when supplied with full data ($p=0.002$). Data Scarcity: No significant performance difference was found under limited-data conditions ($p>0.05$). Efficiency: Riemannian ML architectures train 2–10× faster than Deep Learning models. Optimal Choice: EEGNet provides the best trade-off between accuracy and computational efficiency. Repository Contents src/ : Python source code covering preprocessing, training, evaluation, and visualization. results/ : Complete benchmark results (all_results.csv) and generated figures. notebooks/ : Google Colab-ready reproduction notebooks. environment.yml / requirements.txt : Configuration files for setting up the environment. How to Reproduce Simply open notebooks/reproduce_results.ipynb in Google Colab and run all the cells. Affiliation Department of Informatics, Universitas Logistik dan Bisnis Internasional (ULBI), Bandung, Indonesia.