Selective Ensemble Based on Extreme Learning Machine for Sensor-Based Human Activity Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Sensor-based human activity recognition (HAR) is crucial for health, fitness, and sports applications.
• Existing ensemble methods for HAR can be improved in terms of performance, generalizability, and diversity.

Purpose of the Study:

• To propose a novel HAR approach, DMGSOSEN, utilizing ensemble learning with differentiated extreme learning machines (ELMs).
• To enhance HAR system performance by optimizing classifier diversity and accuracy through selective ensemble learning.

Main Methods:

• Trained multiple base ELMs using bootstrap sampling.
• Pre-pruned the ELM pool by calculating pairwise diversity measures to eliminate similar classifiers.
• Employed glowworm swarm optimization (GSO) to select an optimal sub-ensemble.
• Combined results using majority voting.

Main Results:

• Achieved 96.7% accuracy and high F1 score from wrist data.
• Demonstrated superior performance compared to traditional ensemble algorithms (Bagging, Adaboost) and pruning methods.
• Required fewer base classifiers than existing approaches.

Conclusions:

• The proposed DMGSOSEN method offers improved accuracy and efficiency for sensor-based HAR.
• Balancing diversity and accuracy through selective ensemble learning is effective.
• This approach advances HAR capabilities in various applied fields.