Abstract:

Function perturbation was applied, on the basis of the researches on feature selection for high\|dimensional data, to integrating the results of four different feature selectors in order to get a subset which has good classification accuracy and stability. The former algorithm and the new algorithm were compared on five gene datasets. The experimental results demonstrate that the new algorithm  is able to make the effects of different feature selectors complementary, and thus it effectively improves the stability of feature selection and has a good classification accuracy.

Key words:  high-dimensional data, feature selection, stability, function perturbation, ensemble learning