A Ranking Based Multi-View Method for Positive and Unlabeled Graph Classification

Graph classification becomes an active research problem in these years, since it can deal with the situation where objects contain structure and rich content information. Most of the previous works focus on graph classification by assuming both positive graphs and negative graphs are available. However, in some case of real-world applications, we always collect the positive graphs and a number of the unlabeled graphs, which is referred as the positive and unlabeled graph learning. Moreover, existing graph classification methods are limited to the case where the graph is described from one perspective. In order to address these problems, this paper proposes a new approach, called multi-view positive and unlabeled graph classification (MVPUG). It combines the strategy of cost-sensitive by introducing similarity weight of graphs, which can control the preference of the penalty for different graphs. And it incorporates the different representations of the graph, which exploits the consensus principle and the complementarity principle among different views of graphs. Extensive experiments on real life datasets have shown that our proposed MVPUG can achieve a better performance for multi-view positive and unlabeled graph classification in comparison to the state-of-the-art graph classification methods.