Investigating the Robustness and Stability to Noisy Data of a Dynamic Feature Selection Method

The curse of dimensionality is one of the major problems faced by machine learning researchers. If we consider the fast growing of complex data in real world scenarios, feature selection (FS) becomes a imperative step for many application domains to reduce both data complexity and computing time. Based on that, several studies have been developed in order to create efficient FS methods that performs this task. However, a bad selection of one single criterion to evaluate the attribute importance and the arbitrary choice of the number of features usually leads to a poor analysis. On the other hand, recent studies have successfully created models to select features considering the particularities of the data, known as dynamic feature selection. In this paper, we evaluate one of this successful methods, called pareto front based dynamic feature selection (PF-DFS), to test its stability and robustness in noisy data. We used 15 artificial and real world data with additional noise data. Results shown that the PF-DFS is more stable to noisy scenarios than existing feature selection methods.