Developing Effective Classification Techniques Using Hybrid Feature Fusion and Attention-Guided Classification for Improving Interpretability and Explainability of Human Activity Recognition Models

Abstract

HAR models have shown excellent performance using deep learning approaches; nevertheless, the black box characteristics of such approaches make it hard to deploy and trust such type of models in low safety systems and in the real world. This paper proposes an interpretable and explainable framework for HAR using the combination of improving classification techniques along with transparent representation of features and explanation techniques. The proposed framework uses hybrid feature representation, attention-based classifier, and decision layer based on rules for achieving a superior-performance model. Mathematical formulation of the proposed classifiers, interpretability measures, and explanation measures is provided in this paper. Extensive practical evaluation on the benchmark datasets (UCI HAR, WISDM, and PAMAP2) shows that the method under proposal provides an accurate model while improving its interpretability up to 6-9%.