Effective Sensor Selection for Human Activity Recognition via Shapley Value

Real-time human activity recognition is facing an ever-growing need for efficient sensor setup. Identifying a minimal sensor configuration can lead to cost savings and less intrusive equipment, ultimately improving the quality of the collected data. In this study, we introduce and assess a sensor selection approach that ranks sensors based on their relevance in human activity recognition (HAR) tasks. Our methodology utilizes the Shapley value - a widely adopted metric inspired by game theory - of sensor measurements to determine the importance of each sensor. To validate our approach, we assess the impact of sensor removal on the accuracy of XGBoost tree models, which are trained on a publicly available HAR dataset. Our experiments indicate that Shapley-based sensor ranking achieves a favorable cost-accuracy tradeoff allowing for a reduction by more than 50% in the number of exploited sensors without significantly affecting accuracy.