Neural Network Ensembles for Sensor-Based Human Activity Recognition Within Smart Environments.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Data-driven approaches to human activity recognition (HAR) require high-quality datasets, which are currently scarce for activities of daily living (ADLs) in smart environments.
• Existing HAR datasets often lack the scale, quality, and accurate annotations necessary for robust data-driven models.

Purpose of the Study:

• To improve the quality of an existing HAR dataset for enhanced data-driven HAR.
• To propose a novel ensemble of neural networks for recognizing ADLs in smart home settings.

Main Methods:

• A homogeneous ensemble neural network was developed, integrating four base models using a support function fusion method.
• Data distribution at a class level was explored to minimize conflicts between base models.
• Conflict resolution techniques were investigated, including a differential value approach for conflicting predictions.

Main Results:

• Distributing data at a class level significantly reduced inter-model conflicts, improving performance before explicit conflict resolution.
• The best HAR performance achieved was 80.39%.
• This optimal performance was obtained by combining class-level data distribution with a conflict resolution strategy based on prediction differentials.

Conclusions:

• The proposed ensemble of neural networks, combined with strategic data handling and conflict resolution, effectively enhances HAR performance in smart environments.
• Improving dataset quality and employing class-level data distribution are crucial for successful data-driven HAR.
• The developed conflict resolution method proved effective in maximizing classifier accuracy.