Research Of Brain-network-oriented Feature Selection Method And Its Application

With the increase of feature dimension,in field of machine learning,learning algorithms often have the problem of high dimension with small size of subjects.Feature selection,which can simplify the model and improve the performance of learning algorithms,has been widely applied to machine learning and image analysis.Recently,feature selection methods have been applied to analysis of complex structured data,such as brain connectivity network.However,these studies usually use typical feature selection methods in machine learning,ignoring the intrinsic characteristics of brain connectivity networks,such as the topological properties of the network and the distribution information of network data,which could decrease the performance of brain network analysis.Meanwhile,the brain connectivity network is usually weighted networks.For example,the functional connectivity network based on functional magnetic resonance imaging?fMRI?is a full-connected weighted network.In order to characterize the topology of network,it is necessary to threshold the brain connectivity network using a predefined value.However,there is no golden standard to determine the optimal threshold.On the other hand,brain connectivity networks with different thresholds exhibit different topological properties,which may convey complementary information for further improving the performance of brain network analysis.Based on this background,this paper mainly focus on the research of brain-network-oriented feature selection.The specific work is as follows:?1?In brain network analysis,studies usually extract some local network measurements as features for subsequent feature selection and classification,thus ignoring the intrinsic network topology information and thus leading to the decrease of performance of brain network classification.To address this problem,this paper proposes a graph-kernel feature selection method?called gk-SFS?for analysis of brain networks based on fMRI data.The proposed gk-SFS method not only preserves the topological properties of the brain networks,but also considers the distribution information of the network data.Specifically,the proposed gk-SFS method first introduces a graph-kernel?the kernel constructed on the graph?to calculate the similarity of brain networks,and embeds it into a Laplacian regularization item.Secondly,a L₁-norm sparse item is used to ensure that only a small number of features can be selected.The experimental results on two real fMRI datasets demonstrate that the proposed gk-SFS method can achieve better classification performance than state-of-the-art feature selection methods.?2?In order to take full advantage of the network topological properties and the complementary information conveyed by thresholded networks with different thresholds,this paper extends the proposed gk-SFS method from single threshold to multiple threshold,and thus proposes a graph-kernel-based multi-task feature selection method?called gk-MTFS?.The performance of network analysis could be improved by exploring the complementary information of thresholded network with different thresholds.Specifically,the proposed gk-MTFS explore the complementary information conveyed by thresholded networks with different thresholds by using multi-task learning,and use L_2,1-norm to jointly select the same number of discriminative features from different tasks,and also using graph-kernel-based Laplacian regularization item to preserve the network topological properties and distribution information of network data.Results on two real fMRI datasets demonstrate that the proposed gk-MTFS method can further improve the performance of brain network classification.