The Application Of Feature Selection Algorithm Based On Randomized Structural Sparsity Optimization On Functional Magnetic Resonance Imaging Data Of Cognitive Activities

Cognitive science is a advanced interdisciplinary of exploring the brain and mind working mechanism.The research on cognitive functional magnetic resonance imaging(fMRI)data of cognitive activities can reveal the processing mechanism of cognitive behavior in the brain,and explain and predict human cognitive behavior.However,in the analysis of cognitive fMRI data,the researchers usually only care about the classification effect of the constructed classifier,while ignoring the stability of the optimal feature subset selected by feature selection.But in fact,the fMRI data has the characteristics of high dimension and small sample,so it is prone to the instability of the feature subset,leading to unreliable feature selection results.Therefore,in the process of feature learning of fMRI data,the stability of feature selection is even more important than the classification performance.This thesis uses feature selection algorithm based on the randomized structural sparsity optimization,adding constraint subsampling into the stability selection process,and using the local correlation of voxels in the fMRI data of cognitive activities as a priori structural information.It can maintain a low false negative level while controlling false positives.On the basis of this,we apply the randomized structural sparsity optimization algorithm to two kinds of feature selection problems of cognitive activities:(1)The problem of feature selection based on classification accuracy,in other word,is to improve the classification performance by improving the feature selection method.In this thesis,we study this kind of problem by the experiment of analyzing human emotion recognition mechanism by multi voxel model.Experiments adopt a variety of feature selection methods for feature selection and classification of facial emotion data.The results show that,compared with other methods,the randomized structural sparsity optimization algorithm can get the highest classification accuracy,and can also reveal the brain activation related to emotion recognition better.(2)The problem of feature selection based on the accuracy of voxel selection,that is to say,the feature selection is more important than the classification accuracy.In such problems,using only a small fraction of the true discriminative region to do classification,the classification accuracy can easily reach very high or even 100%.At this point,there is no point in paying attention to the accuracy of classification.What should be paid attention to is feature selection.In this article,we study this kind of problem through the multi-voxel pattern analysis of how human brain recognize neutral face and how human brain recognize happy face.And experiments adopt a variety of feature selection methods for feature selection of face data.The results show that,compared with other methods,the randomized structural sparsity optimization algorithm can detect the activated brain areas related to face processing more comprehensively,the brain regions are more accurate and compact,the potential false positives are less,and brain regions are of strong interpretability.