Predicting Emotion with Biosignals: A Comparison of Classification and Regression Models for Estimating Valence and...

Area of Science:

• Physiological computing
• Affective computing
• Machine learning for biosignal analysis

Background:

• Emotion recognition is crucial for human-computer interaction.
• Biosignals offer objective measures of emotional states.
• Previous models have limitations in accuracy and generalizability.

Purpose of the Study:

• To predict emotional valence and arousal using wrist-worn biosensor data.
• To compare the performance of various machine learning models for emotion prediction.
• To investigate the impact of normalization methods and sensor selection on prediction accuracy.

Main Methods:

• Collected biosignals from wrist-worn sensors and self-reported emotional states via questionnaires.
• Implemented and compared diverse classification and regression models, including Long Short-Term Memory (LSTM) networks.
• Evaluated different data normalization techniques and the effect of using subsets of biosignals.

Main Results:

• Regression models outperformed classification models, with LSTM regression yielding the best performance.
• The 'baseline reduction' normalization method proved most effective.
• An LSTM regression model with baseline reduction achieved high accuracy for valence (MSE=0.43, R2=0.71) and arousal (MSE=0.59, R2=0.81).
• Optimal models were sometimes achieved using only a subset of available biosignals, varying by individual.

Conclusions:

• LSTM-based regression models are highly effective for predicting valence and arousal from biosignals.
• The baseline reduction normalization technique enhances emotion prediction accuracy.
• Feature selection, utilizing fewer biosignals, can lead to robust and personalized emotion recognition models.