Feature selection for multimodal emotion recognition in the arousal-valence space

Area of Science:

• Multimodal emotion recognition
• Affective computing
• Human-computer interaction

Background:

• Emotion recognition research traditionally relies on facial expressions.
• Integrating multiple biosignals shows promise for improving emotion classification accuracy.
• Identifying the most relevant biosignals remains an open challenge.

Purpose of the Study:

• To identify the most relevant biosignals for accurate emotion recognition.
• To determine which biosignals contribute most to classifying arousal and valence.
• To reduce feature dimensionality while maintaining classification performance.

Main Methods:

• Utilized Recursive Feature Elimination (RFE) for feature selection.
• Employed a multimodal database with arousal and valence annotations.
• Extracted diverse features from physiological, neurophysiological, and video signals.

Main Results:

• Recursive Feature Elimination successfully identified a subset of relevant features.
• Classification accuracy was preserved even after eliminating several features for 2 and 3 class setups.
• Achieved up to 70% accuracy for arousal and 60% for valence using a reduced feature set.
• Galvanic Skin Response (GSR) proved relevant for arousal, while Electroencephalogram (EEG) was relevant for valence.

Conclusions:

• A reduced set of biosignals can effectively support emotion recognition.
• Galvanic Skin Response (GSR) and Electroencephalogram (EEG) are critical for specific emotion dimensions (arousal and valence, respectively).
• Feature selection techniques like RFE are valuable for optimizing multimodal emotion recognition systems.