Emotion Recognition Research using XAI and Fair AI

Overview

Designed GRU-based emotion classification system for multi-class emotion recognition using physiological biosignals including ECG (electrocardiogram), EMG (electromyography), GSR (galvanic skin response), respiration rate, and BVP (blood volume pulse), achieving 95% classification accuracy.

Key Contributions

• Implemented SHAP (Shapley Additive Explanations) and LIME (Local Interpretable Model-agnostic Explanations) for comprehensive feature contribution analysis
• Conducted comparative feature importance analysis across ECG, EMG, GSR, respiration, and BVP signals using Shapley value calculations and game theory-based approaches
• Implemented fair ML strategies including Learning Fair Representations (LFR), Disparate Impact Remover (DIR), and Reweighing techniques to systematically reduce demographic bias by 30% while maintaining high classification performance
• Investigated fairness and bias using multiple evaluation metrics including Statistical Parity Difference (SPD), Equal Opportunity Difference (EOD), and Theil Index (TI)
• Enhanced model transparency by providing explainable predictions with visual feature attribution maps

Technologies

GRU, SHAP, LIME, Fairness in ML, Bias Mitigation, Deep Learning, Explainable AI