Representation Learning for Interpersonal and Multimodal Behavior Dynamics: A Multiview Extension of Latent Change...

Area of Science:

• Computational behavior analysis
• Psychotherapy research
• Stochastic systems modeling

Background:

• Analyzing multimodal and interpersonal behavior dynamics is crucial for understanding complex interactions.
• Conventional methods often fail to capture causal relationships or precise behavioral patterns.
• Individualized tracking is vital in applications like psychotherapy for assessing patient mental states.

Purpose of the Study:

• To present a novel approach for learning multimodal and interpersonal representations of behavior dynamics.
• To enable the concurrent capture of inter-modal and interpersonal behavior dynamics and identify directional relationships.
• To enhance the interpretability and predictive performance of computational behavior analysis.

Main Methods:

• Developed a multiview extension of latent change score models.
• Applied the approach to therapist-client interactions to study collaborative dynamics.
• Integrated the model with probabilistic classifiers, such as Gaussian process models, by explicitly modeling uncertainty.

Main Results:

• The novel approach demonstrated improved performance over conventional methods relying on summary statistics or correlational metrics.
• The multiview extension successfully captured both inter-modal and interpersonal behavior dynamics and their directional relationships.
• Integration with probabilistic classifiers further enhanced predictive performance while maintaining interpretability.

Conclusions:

• The proposed multiview latent change score model offers a powerful and interpretable method for analyzing complex behavior dynamics.
• The approach provides deeper insights into interpersonal relationships, particularly in therapeutic contexts.
• Further exploration of stochastic systems in computational behavior models is warranted.