A novel multigranularity feature-selection method based on neighborhood mutual information and its application in autistic patient identification

Abstract

The high dimensionality and small sample of functional magnetic resonance imaging (fMRI) data is the big challenge for machine learning application in identification of mental disorders from fMRI images. Feature selection provides an effective method to select the task related fMRI features and removing the redundant ones. The existing feature-selection methods improved the performance of machine learning model by selecting the discriminative features under the limitation of weak correlation among candidate features. However, the strong correlation among fMRI features and its actual influence on classification performance is less considered. Herein, a novel multigranularity feature-selection method was proposed, which considers both the feature’s discrimination and the correlation between features at the same time. Firstly, k-means clustering was used to divide fMRI samples into subgroup reducing the potential heterogeneity within subgroups. Second, a new weight proportional to features’ correlation and inversely proportional to the discrimination was used to create minimum spanning trees representing the fMRI feature space. Third, the impact of the correlation among features on the classification was further examined from optimistic and pessimistic perspectives with multigranularity information. The experimental results on fMRI data from ABIDE database show that our method not only reduced the feature redundancy but also is superior to a variety of competing feature-selection methods in autism spectrum disorder (ASD) recognition.