Group based techniques for stable feature selection

Feature selection is an intensely active research area in data mining and machine learning. Much of this research has focused on selecting subsets of features that improve predictive accuracy of classification models, but the stability of these feature subsets has been little studied. Feature selection stability is critical in domains where identifying intrinsically important features is just as necessary as making an accurate prediction. One example of this is when biologists use feature selection on high-dimensional and small sample genomic microarray data as a means to identify biomarkers that respond to the presence or absence of a disease. In this work we study the causes of feature selection instability, and propose two feature group based feature selection frameworks that enhance the stability while not sacrificing the predictive power of the resulting feature subsets. Our experimental study shows that two algorithms developed under the proposed frameworks are effective at producing more stable feature selection results with comparable accuracy to state-of-the-art feature selection algorithms.