Machine Learning Techniques for Sensor-Based Human Activity Recognition with Data Heterogeneity-A Review.

Area of Science:

• Ubiquitous Computing
• Machine Learning
• Sensor Data Analysis

Background:

• Human Activity Recognition (HAR) is vital for analyzing behaviors using multi-dimensional sensor data.
• Current HAR research often overlooks data distribution variations across datasets, limiting real-world applicability.
• Data heterogeneity in sensor data poses a significant challenge to HAR model generalization and efficiency.

Purpose of the Study:

• To review machine learning approaches for addressing data heterogeneity in Human Activity Recognition (HAR).
• To categorize types of data heterogeneity encountered in HAR.
• To identify suitable machine learning methods and datasets for heterogeneous HAR.

Main Methods:

• Categorization of data heterogeneity types in sensor-based HAR.
• Identification and application of machine learning methods tailored to specific heterogeneity challenges.
• Systematic review of existing HAR datasets and their characteristics.
• Discussion of future research directions and open challenges.

Main Results:

• Identified distinct types of data heterogeneity impacting HAR performance.
• Highlighted machine learning techniques effective in mitigating these heterogeneity issues.
• Provided a summary of relevant HAR datasets, noting their distributional properties.
• Emphasized the benefits of addressing heterogeneity, including improved accuracy and reduced computational load.

Conclusions:

• Machine learning offers viable solutions to overcome data heterogeneity in HAR.
• Addressing data distribution variations is key to developing robust, personalized, and efficient HAR systems.
• Further research is needed to explore advanced adaptive and federated learning approaches for HAR.