Using Domain Knowledge for Interpretable and Competitive Multi-Class Human Activity Recognition.

Area of Science:

• Machine Learning
• Human Activity Recognition (HAR)
• Signal Processing

Background:

• Human Activity Recognition (HAR) is a growing machine learning application.
• Multi-class HAR performance is sensitive to problem decomposition, yet research is limited.
• Existing studies lack empirical comparisons of decomposition methods in HAR.

Purpose of the Study:

• To empirically compare multi-class decomposition methods for HAR.
• To evaluate the performance of machine learning algorithms with different decomposition strategies.
• To investigate the effectiveness of expert hierarchies and their ensembles in HAR.

Main Methods:

• Compared five machine learning algorithms using four multi-class decomposition methods.
• Utilized expert hierarchies (nested dichotomies) and ensembles on a 17-class HAR dataset.
• Extracted features from tri-axial accelerometer and gyroscope signals.
• Evaluated performance on the multi-class problem and two binary classification problems.

Main Results:

• Expert hierarchies compete effectively with the one-vs-all method for HAR.
• Ensembles of expert hierarchies outperform one-vs-all and match one-vs-one in multi-class HAR.
• Ensembles of expert hierarchies show superior performance on binary classification tasks derived from HAR data.

Conclusions:

• Expert hierarchies are a viable alternative to traditional decomposition methods in HAR.
• Ensemble methods based on expert hierarchies offer efficient and effective solutions for multi-class HAR.
• The choice of decomposition strategy significantly impacts HAR system performance.