Metaheuristic-Driven Feature Selection for Human Activity Recognition on KU-HAR Dataset Using XGBoost Classifier.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Human Activity Recognition (HAR) models often rely on complex deep learning approaches.
• There is a need for efficient and accurate HAR methods using readily available sensor data.

Purpose of the Study:

• To develop and evaluate novel HAR frameworks using Extreme Gradient Boosting (XGBoost) enhanced with metaheuristic algorithms.
• To compare the performance of Golden Jackal Optimization-XGBoost (GJO-XGB) and War Strategy Optimization-XGBoost (WARSO-XGB) against traditional classifiers.

Main Methods:

• Utilized the KU-HAR dataset from smartphone accelerometer and gyroscope sensors.
• Extracted 48 mathematical features for HAR.
• Implemented and compared GJO-XGB and WARSO-XGB frameworks using 10-fold cross-validation.

Main Results:

• WARSO-XGB achieved the highest mean accuracy (94.04%), F-score (92.88%), precision (93.47%), and recall (92.40%).
• Both GJO-XGB and WARSO-XGB demonstrated competitive performance, with GJO-XGB showing more stable results.
• WARSO-XGB exhibited lower time complexity (30.84s training, 0.51s testing) compared to GJO-XGB.

Conclusions:

• The proposed WARSO-XGB framework offers a highly accurate and efficient solution for human activity recognition.
• Feature importance analysis using SHAP identified key sensor features contributing to HAR accuracy.
• These XGBoost-enhanced metaheuristic approaches provide a viable alternative to complex deep learning models for HAR.