Affective State Recognition with Convolutional Autoencoders.

Area of Science:

• Affective computing
• Machine learning
• Wearable sensor technology

Background:

• Accurate classification of affective states is crucial for mental health applications.
• Traditional machine learning requires manual feature extraction, which is time-consuming and may not capture complex patterns.
• Deep learning, like Convolutional Neural Networks (CNNs), can automate feature extraction but may overfit small datasets.

Purpose of the Study:

• To develop a robust and generalizable model for classifying affective states using physiological signals from wearable sensors.
• To overcome the limitations of traditional machine learning and standard CNNs in feature extraction and overfitting.
• To enhance the accuracy of affective state recognition through an unsupervised feature learning approach.

Main Methods:

• Utilized unsupervised convolutional autoencoders for automatic feature extraction from time-series physiological data.
• Fed extracted features into a supervised classifier for affective state classification.
• Employed the WESAD (Wearable Stress and Affect Detection) dataset for model training and evaluation.

Main Results:

• Achieved an accuracy increase of nearly 3% over traditional CNN models using all physiological data.
• Demonstrated a 2% accuracy improvement using chest-only physiological data.
• Showcased an 8% accuracy gain using wrist-only physiological data for classifying neutral, stress, and amusement states.

Conclusions:

• The proposed unsupervised convolutional autoencoder approach offers a more robust and accurate method for affective state classification compared to standard CNNs.
• This method effectively extracts relevant features from physiological time-series data, improving generalization and performance.
• The developed system has significant potential for clinical applications, including patient monitoring and cognitive therapy.