X-CHAR: A Concept-based Explainable Complex Human Activity Recognition Model.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning models for human activity recognition (HAR) are crucial for safety-critical applications but often lack interpretability.
• Existing methods for explainable HAR require extensive, precise low-level activity annotations, increasing expert workload and error potential.
• Bridging the gap between high-performance deep learning and trustworthy, interpretable HAR remains a significant challenge.

Purpose of the Study:

• Introduce X-CHAR, an eXplainable Complex Human Activity Recognition model.
• Develop a deep learning approach for HAR that provides human-understandable explanations without precise low-level activity annotations.
• Reduce the burden on domain experts while maintaining robust prediction accuracy.

Main Methods:

• X-CHAR models complex activities as sequences of high-level concepts.
• Utilizes Connectionist Temporal Classification (CTC) loss to handle sequence information without requiring exact start/end times for low-level annotations.
• Generates explanations in the form of concept sequences and counterfactual examples.

Main Results:

• X-CHAR achieves robust performance comparable to baseline end-to-end deep learning models for time series data.
• The model successfully provides explanations in the form of understandable, high-level concepts.
• A mechanical Turk study confirmed that X-CHAR's explanations are more understandable than those from existing methods.

Conclusions:

• X-CHAR offers a viable solution for explainable complex HAR, reducing developer burden and enhancing model trust.
• The approach effectively balances prediction accuracy with interpretable explanations for time series data.
• This work advances the field of interpretable AI in safety-critical HAR applications.