Sneaky emotions: impact of data partitions in affective computing experiments with brain-computer interfacing.

Area of Science:

• Neuroscience and Machine Learning
• Computational Neuroscience
• Affective Computing

Background:

• Brain-Computer Interfaces (BCI) show potential for emotion recognition using machine learning (ML).
• The impact of data partitioning in training/test splits on BCI research findings is often overlooked.
• This oversight complicates attributing results to model improvements versus data splitting artifacts.

Purpose of the Study:

• To introduce and utilize the 'data transfer rate' concept to analyze data partitioning effects in BCI.
• To investigate how different data splitting strategies influence emotion recognition performance using EEG signals.
• To provide guidelines for managing and reporting brain data partitioning in BCI research.

Main Methods:

• Introduced the 'data transfer rate' construct, quantifying overlap between training and test data.
• Examined emotion recognition from electroencephalogram (EEG) signals in videos under three data partitioning conditions: subject-independent, video-independent, and time-based splits.
• Evaluated classification accuracy changes across different partitioning strategies.

Main Results:

• Model performance varied significantly (50%-90% accuracy) based on data partitioning methods.
• Subject-independent (affective decoding) performance remained near baseline unless test subjects' data were included in training.
• Video-independent (affective annotation) performance improved when training and test data shared the same subject, even with different videos.
• Later EEG signal segments were more discriminative, but the sheer number of data points was the most critical factor.

Conclusions:

• Data partitioning is a critical factor influencing BCI performance in emotion recognition tasks.
• Specific partitioning strategies are more suitable for different BCI applications (e.g., affective decoding vs. annotation).
• Standardized reporting of data partitioning methods is essential for reproducibility and accurate interpretation of BCI research findings.