Emotional Recognition in Gaming Using EEG Signals

Problem Statement

Video games elicit complex emotional responses. Understanding these responses can enhance user experience, enable adaptive game design, and support mental health applications.

1. EEG-based emotion recognition provides a non-invasive method to measure brain activity in real-time and classify emotional states. Accurate detection of positive and negative emotions allows:
2. Personalizing gaming experiences
3. Identifying stress or boredom triggers
4. Integrating emotion-adaptive mechanics

This project applies machine learning to EEG (Electroencephalography) data from 27 participants to classify emotional states.
The task was simplified into binary classification — grouping Calm, Satisfaction, and Funny as positive emotions, and Boredom and Horrible as negative emotions.

We tested SVM, Logistic Regression, and Random Forest, with Random Forest achieving the best performance.

Key Results

1. Achieved 62.9% accuracy in binary emotion classification using Random Forest
2. Random Forest outperformed SVM and Logistic Regression for EEG-based emotion recognition
3. Feature importance analysis showed distributed contributions across multiple brain regions rather than concentrated in a few channels

Methodologies

1. Collected EEG data from 27 participants playing 4 games (5 min each)
2. Preprocessed EEG signals: averaged channels, standardized features, and converted to binary labels
3. Split data into 80% training and 20% testing
4. Trained SVM, Logistic Regression, and Random Forest models
5. Evaluated models using accuracy and confusion matrices
6. Analyzed feature importance to understand brain-region contributions

Data Sources

• Source: Cleaned EEG CSV files from 27 participants, 4 games each (5 minutes per game)
• Features: Averaged EEG channels by brain region (frontal, temporal, parietal, occipital) plus mean and variance
• Labels: Binary emotion labels — Positive (Calm, Funny, Satisfaction) and Negative (Boredom, Horrible)

Kaggle Datasets: Link to Kaggle Dataset

Technologies Used

• Python
• pandas, numpy, scikit-learn
• Matplotlib, Seaborn (for visualizations)

Models Tested

• SVM — Moderate performance, struggled with high-dimensional space
• Logistic Regression — Simple and interpretable, but underperformed
• Random Forest — Best performance, handled non-linear relationships well
• Feature importance: Spread across multiple brain regions rather than concentrated in a few channels

Results

• Accuracy: 62.9%
• Confusion Matrix: [[275405 177988] [113154 218837]]

Authors

This project was completed in collaboration with:

• Veda Manchana ([email protected])
• Marzia Tahsin ([email protected])