Comparison of wrapper and filter feature selection algorithms on human activity recognition

Feature selection is an increasingly important part of machine learning. The purpose of feature selection is dimension reduction in a large multi-dimensional data set and it can be the key step of successful knowledge discovery in those problems where the number of features is large. This research area has huge practical significance because it accelerates decisions and improves performance. The requirements of specific applications in different kinds of research areas have led to the development of new feature selection techniques with different properties. In the last few decades, several feature selection algorithms have been proposed with their particular advantages and disadvantages. Despite of the intensive research and the large amount of works, the different kinds of feature selection algorithms have not been tested yet in the human activity recognition problem. It was the main motivation of our work and this paper tries to fill this gap. Therefore, in this article we present a conceptually simple naive Bayesian wrapper feature selection method and compare it with some widely used filter feature selection algorithms. The result of this work demonstrates that, the wrapper technique outperforms filter algorithms in this type of problem. In addition, this paper shows an example, when the classifier dependency of a wrapper method do not visible.