Personalization of Affective Models Using Classical Machine Learning: A Feasibility Study.

Area of Science:

• Affective computing
• Machine learning
• Human-computer interaction

Background:

• Traditional emotion recognition models often use a one-size-fits-all approach.
• Personalization in affective computing can lead to more accurate and nuanced emotion detection.

Purpose of the Study:

• To investigate the effectiveness of personalized machine learning models for emotion recognition.
• To compare personalized models against generalized models using physiological and video data.

Main Methods:

• Trained personalized k-nearest neighbors, random forests, and dense neural networks on 51 features from the Emognition dataset.
• Ensured temporal separation between training and testing data for each subject.
• Compared personalized model performance (F1 scores) against generalized models trained on all subjects.

Main Results:

• Personalized models achieved higher mean F1 scores (e.g., 92.66% for random forests) compared to generalized models (e.g., 91.78%).
• Personalized models outperformed generalized models for 7 out of 10 subjects.
• Principal Component Analysis (PCA) indicated that low intra-subject data variability hinders personalized model performance.

Conclusions:

• Personalized machine learning models show significant potential for improving emotion recognition accuracy.
• The effectiveness of personalized models is contingent on sufficient within-subject data variation.
• Further research is needed to address challenges in implementing personalized affective computing.