Abstract
Accurate human emotion recognition is a fundamental challenge in affective computing. This study presents a methodology for classifying emotional states using Galvanic Skin Response (GSR) signals. The approach integrates wavelet denoising to isolate emotional signatures and a covariance-based feature selection technique to reduce redundancy. By selecting the 15 most independent features from an initial set of 30, the researchers optimized a Support Vector Machine (SVM) classifier. Experimental results on a custom dataset of five emotional states (happiness, grief, anger, fear, and calm) demonstrate that the feature selection process increased test accuracy from 46.67% to 66.67%, with fear achieving a 100% recognition rate.
Introduction
Emotion recognition is increasingly vital for human-computer interaction, yet physiological variability makes it difficult to implement reliably. Galvanic Skin Response (GSR), which measures skin conductivity changes driven by sweat gland activity, serves as a sensitive indicator of emotional arousal. However, raw GSR data is inherently noisy and high-dimensional. When researchers extract numerous statistical features, many are highly correlated, leading to the "curse of dimensionality" and model overfitting. This paper addresses these issues by proposing a streamlined pipeline for signal processing and feature optimization.
Methodology
The researchers developed a multi-stage pipeline to transform raw physiological data into emotional classifications.
- Signal Acquisition & Denoising: GSR signals were captured using the e-Health Sensor Platform V2.0. To eliminate high-frequency noise and baseline drift, the team employed wavelet denoising with the
db5function, effectively creating a band-pass filter (0.08–0.2Hz). - Normalization: To mitigate individual physiological differences, data was normalized by comparing active emotional responses against a baseline "calm" state for each subject.
- Feature Extraction & Selection: Initially, 30 features were extracted across time and frequency domains. To combat redundancy, a covariance-based selection method was applied. By identifying and retaining only the 15 features with the lowest mutual correlation, the researchers ensured the SVM received a diverse and independent information set.
- Classification: An SVM with a Radial Basis Function (RBF) kernel was used to classify signals into five categories: happiness, grief, anger, fear, and calm.
Experimental Setup
The study utilized a custom database involving junior students at Jilin University. Emotions were elicited using four movie fragments (4–5 minutes each), interspersed with recovery periods to establish a calm baseline. The final dataset comprised 257 emotion files: Happiness (57), Grief (51), Fear (47), Anger (43), and Calm (59).
Results and Discussion
The core finding of this research is the dramatic impact of feature selection on model performance. While the 30-feature model showed high training accuracy, it failed to generalize to new data.
- Accuracy Improvement: The 15-feature model achieved a test accuracy of 66.67%, compared to only 46.67% for the 30-feature model. This confirms that redundant features acted as noise, causing the model to memorize training data rather than learn underlying patterns.
- Distinctive Emotions: Notably, "Fear" reached a 100% recognition rate in the optimized model, suggesting its physiological signature in GSR is uniquely identifiable.
- The "Less is More" Principle: Reducing the feature set by half nearly doubled the testing performance, highlighting the importance of minimizing covariance in physiological signal processing.
Conclusion
This study demonstrates that covariance-based feature selection is a powerful tool for improving GSR-based emotion recognition. By focusing on independent features and utilizing wavelet denoising, the researchers created a more robust SVM classifier. Future work will focus on expanding demographic diversity and adapting the model for real-time applications where motion artifacts present additional challenges.
Summary updated