Dense Subgraph Mining Based Brain Network Classification With Applications

The brain diseases not only threat the life of patients, but also bring heavy burden on society, so the researchers pay more attention on brain diseases. With the development of neuroimaging technology,the researchers applied neuroimaging technology on brain diseases and obtained great progress. The researchers mined brain functional network from neuroimaging , then use machine learning and data mining to mine features from brain networks which is applied on brain network classification. The paper research brain diseases based on brain network, the main works summarized as follows:First part, we proposed method for brain network classification based on dense subgraph parti-tion. This part utilized the dense subgraph to reflect the topological pattern with close connection in human brain. Specifically, we convert the brain network to edge-dual graph which is segmented to dense subgraphs with decreasing density. Then we choose several dense subgraphs with highest den-sity to reconstruct the brain network. Finally we use graph kernel to measure the similarity between reconstructed networks, construct graph kernel matrix and use SVM to classify. The experiment results shows the proposed method can improve the classification performance and can mine regions related to brain diseases.Second part, we combined the idea of dense subgraph and frequent subgraph and proposed method for brain network classification based on frequent dense subgraph. This part is based on the fact that there exists significant differences between patients and the normals, so the frequent dense subgraph-s mined from different dataset exists significant differences which can be applied on classification.Specifically, we mined frequent dense subgraphs from patients and normal dataset and computed the discriminative ability of each subgraph. Then we choose the subgraphs with highest discriminative abil-ity and construct indicating matrix as feature matrix. Finally, we use SVM to classify. The experiment results indicates the proposed method can improve the accuracy and can mine topological pattern related to brain diseases.