Robust Health Score Prediction from Pyro-Sensor Activity Data based on Greedy Feature Selection

Automated activity assessment using IoT/smartphone sensors becomes great popular in ubiquitous computing research community recent year thanks to the enhancement of mobility and IoT sensing. In these researches, owing to the great success of statistical machine learning technique called Lasso, the work offers the interpretability of the model. However, in some sparse feature condition, Lasso as a $l_{1}$ regression method could not give a satisfying result for prediction precision and feature selection. In this paper, we propose a new prediction scheme using greedy feature selection method which is expected to be effective under large scale feature in limited number of dataset. With the help of the new scheme, we could solve the overfitting problem when using $l_{1}$ regression as well as giving satisfying prediction result. Experimental results using longitudinal pyro-sensor dataset of health score of elderly people show that our new scheme offers better interpretability as well as achieves better prediction accuracy compared with Lasso