Feature Selection Methods For Graph Data And Its Applications

As the current hot topic of data mining, machine learning, pattern recognition, Feature selection methods has been widely used in image retrieval, face recognition, text mining, intrusion detection and other fields. But with the development of the times, a variety of study fields has accumulated a large number of graph data, feature selection for graph data become a hot topic. Especially the development of brain network research, need more feature selection methods proposed and improved. The traditional method feature selection for vector data, in particular, has certain limitations when apply to brain network or graph data. This thesis regards subgraph as feature of graph data, the methods proposed can be used to classify the brain network. Our main work is as follows:Firstly, by using the subgraph features method made the representation of brain network need not to be limited to a single brain region, discriminative subgraph features were selected based on the difference of frequency between different classes. Then, graph kernel method was also introduced to measure the similarity between structural data, and also the dimension reduction method based on graph kernel. Our proposed method has been validated on real mild cognitive impairment data sets, the experimental results show that the subgraph feature and the frequency difference based feature selection method can effectively improve the classification accuracy and achieve better interpretability.Secondly, we utilized HSIC criteria which measures features and sample labels dependence method for graph data classification. Frequency difference based feature selection method has good interpretability, but may miss important information for classification. Our proposed method has been validated on real attention deficit hyperactivity disorder data sets,the experimental results show the proposed method can improve the classification performance.In the end, we proposed a brain network classification method based on multi-feature fusion, no matter vector feature or subgraph feature, there will be loss of sample information:vector based feature may lost the topological information between multiple brain region, and subgraph based feature is not sensitive to the change of single brain region. Experimental results show the method have better classification performance under multiple thresholds and the complementary of different types of feature representation for a sample.