Data Valuation Algorithm for Inertial Measurement Unit-Based Human Activity Recognition.

Area of Science:

• Machine Learning
• Biomedical Engineering
• Data Science

Background:

• Human Activity Recognition (HAR) using Inertial Measurement Units (IMUs) is crucial for healthcare and sports.
• Existing data valuation methods often require feature-level input, limiting their applicability.
• Raw IMU data presents challenges due to its multivariate time-series nature.

Purpose of the Study:

• To propose a novel data valuation algorithm for IMU-based HAR data.
• To enable the algorithm to process raw-level inputs by incorporating feature extraction.
• To enhance the reliability and performance of IMU-based HAR models.

Main Methods:

• A meta reinforcement learning-based data valuation algorithm is proposed.
• A feature extraction structure, including a 1D-CNN backbone and a transformer encoder, is integrated.
• A 1D-CNN-based stacking ensemble predictor is utilized for supervised model training.

Main Results:

• The proposed algorithm achieves excellent valuation performance on IMU-based HAR datasets.
• Corrupted data is discovered at a rate exceeding 96% across multiple datasets.
• Classification performance is significantly improved by suppressing low-value data identification.

Conclusions:

• The developed data valuation algorithm effectively handles raw IMU data for HAR.
• The algorithm demonstrates high accuracy in detecting corrupted data.
• Improved classification performance validates the utility of the proposed data valuation approach.