Source Data Selection - public
Description
This repo contains the code for belonging to the paper: Source Data Selection for Brain-Computer Interfaces based on Simple Features by Frida Heskebeck, Carolina Bergeling, and Bo Bernhardsson. The repo has two main scripts data_generation and source_selection. The first script generates the data, i.e. do the transfer learning and evaluate cross subject performance, while the second script predicts which source user is best for each target user and evaluates different methods for souce data selection.
The data used in the paper is included in the source_selection folder.
Visuals
From data_generation script
The following image is generated from the data_generation script.
Figure 1. Transfer learning accuracy sorted on intra subject performance for a selection of users.
From source_selection script
The following images are generated from the source_selection script.
Figure 2. Transfer learning accuracy sorted on column and row sum.
Figure 3. Comparison of different methods for source data selection. TransPerfPred is the suggested method in the paper. Highlighted cells indicate no statistical significance.
Figure 4. Comparison of different methods for source data selection with varying number of selected source users compared to oracle performance. TransPerfPred is the suggested method in the paper.
Installation
The code is run using Python 3.12 with the required packages specified in requirements.txt.
Usage
Run the script data_generation to generate the data and the above shown figure.
Run the script source_selection to analyze the data and generate the above shown images. To use the data generated from data_generation one must move the data to the source_selection/data folder.
Authors and acknowledgment
The code is written by Frida Heskebeck. The research project was supervised by Carolina Bergeling and Bo Bernhardsson.
License
MIT Licence. Basically - Do whatever you like with the code but you are yourself responsible for it.
Project status
The academic paper is to be submitted for publication.