Investigation On PU Graph Classification Based On Multi View Learning

Graph is a kind of abstract data structure,which consists of consists of vertices and links of each vertice pairs.With development of data mining and information science,researchers find that graph has great expressive power,which can model different kinds of data with complex structures in different domains,like DNA and protein in bioinformatics,chemical compound in chemistry,social networks and knowledge graph in computer science.How to classify the graph becomes an active research problem.However,in some case of real-world applications,we always collect the positive graphs and the unlabeled graphs,which is referred as the positive and unlabeled graph learning.Most of the existing work on graph classification always assume that the positive graphs and negative graphs can be collected and are not suitable for the situation of positive and unlabeled graphs,which limits the application of graph classification in real life.In addition,these graph classification methods only describe the graph from one perspective and leads to the lack of classification performance and generalization ability.Aiming at the problems of how to design an appropriate classifier,how to improve classification accuracy targeted at the graph,and how to improve the classification accuracy in case of the PU graph problem,this thesis proposes a new multi view learning method for PU graph classification(MVPUG).Different from the general graph classification method,this method can calculate the score of graphs to obtain the order of prediction accuracy of graphs,which can help researchers to give priority to the graphs with high prediction accuracy.The main contribution of our work can be summarized as follows:(1)We propose a novel model for PU grapp classification.The classify model incorporate a novel constraint so that the model can gain consistent classification results in different views of the graph,which satisfies the principles in multi-view learning.In addition,the model incorporates another constraint such that the classifier can obtain a similar result if the graph is similar,which utilize the similarity between the unlabeled graphs to improve classification performance.(2)We propose a PU graph classification framework,which consists of two parts,learning stage and prediction stage.In the learning stage,it reorganizes the graph set according to the labels of training samples,and uses different graph feature extraction methods to generate the multi-view data of graph.Then,it finds the scoring function of the optimal classification model through cross validation.In the prediction stage,it constructs the multi-view data of the graph that to be classified,then uses the classifier model to calculate the score of the graph,and finally classifies the graph data according to the calculated score.(3)To investigate the performance,we have conducted experiments of our proposed MVPUG method.In the experiments,a number of real datasets are used,and the results show that the proposed MVPUG method performs better and more stable than the existing classification methods.