New prior knowledge based extensions for stable feature selection

In many data sets, there are only hundreds or fewer samples but thousands of features. The relatively small number of samples in high dimensional data results in modest classification performance and feature selection instability. In order to deal with the curse of dimensionality, we propose to investigate research on the effect of integrating background knowledge about some dimensions known to be more relevant, as a means of directing the feature selection process. We propose extensions of three feature selection techniques, two filters and a wrapper, by incorporating prior knowledge in the search procedure of the best features. We study the effect of these extensions on the classification performance and the stability of the feature selection. We experimentally test and compare our proposed approaches with their original versions, which do not integrate prior knowledge, over three high-dimensional datasets. The results show that our proposed techniques outperform other methods in terms of stability of feature selection but also in classification performance in most cases.