Personalized Clustering for Emotion Recognition Improvement.

Area of Science:

• Affective computing
• Artificial intelligence
• Human-computer interaction

Background:

• Emotion recognition via AI and smart sensing shows promise but requires improvements for real-world safety applications.
• Current systems need to be fast, discrete, and real-time, especially for sensitive areas like violence detection and mental health.
• General AI models are insufficient for multi-user protection due to individual differences in emotional responses.

Purpose of the Study:

• To develop customized, lightweight AI models for emotion recognition tailored to clusters of individuals with similar physiological responses.
• To address the need for semi-personalized models in real-life applications, enhancing safety and well-being systems.
• To present a methodology for clustering users and creating adaptable AI models that can incorporate new, unlabeled data.

Main Methods:

• Clustering of subjects based on labeled physiological data and emotional responses.
• Development of individualized AI models for each identified cluster.
• Implementation of a method for enrolling new users with unlabeled data into existing models.
• Continuous updating of AI models with new incoming data.

Main Results:

• Demonstrated a 5% improvement in accuracy compared to a general baseline model.
• Achieved a 4% increase in F1 score over the baseline model.
• Reported a significant reduction in model variability, ranging from 32% to 58%.

Conclusions:

• The proposed methodology for semi-personalized emotion recognition models enhances performance and reduces variability.
• This approach is adaptable for expert systems dealing with unlabeled data and diverse user profiles.
• The findings support the development of more effective and personalized AI solutions for safety and mental health monitoring.