Improving Noise Robustness of Single Sensor Data in Human Activity Recognition With UMAP and Additional Data
Quynh Nguyen Phuong VU, Paula Lago, Sozo Inoue · 2022
In this paper, we propose a method for human activity recognition that can transfer information from multiple sensors to a single sensor with improving noise robustness by using a topological feature representation created by UMAP. With clean data, UMAP improves as much as 12% in macro F1-score compared to the Feature Agglomeration approach. With noisy data, UMAP is more stable, with only a 0.25% decrease in average macro F1-Score. By contrast, with Feature Agglomeration and Traditional approaches, the average macro F1-scores decline by 2.75% and 5.75%, respectively.