Boundary Oversampling Based Graph Node Imbalance Classification Algorithm.

In the real world, financial fraud detection and disease diagnosis are typical instances of graph imbalanced problems. Graph neural networks based on oversampling are among the commonly employed methods to address such issues. However, this approach encounters challenges in ensuring the diversity of generated boundary samples, which can lead to a reduction in classification performance. This paper introduces a graph node imbalanced classification algorithm based on borderline oversampling (ImBS) to enhance the diversity of generated samples. Firstly, ImBS selects high-confidence samples from each class as sampling anchors using a two-layer graph neural classification network, enhancing the representativeness of the anchors. Next, to make the distribution of generated samples more reasonable, this paper utilizes the obtained confusion matrix from the previous step to calculate the distribution ratio of misclassified instances in the minority class. Based on this distribution ratio, an adaptive computation of the number of generated samples among different classes is proposed. Building upon this, a hybrid oversampling method based on anchors is introduced. It oversamples boundary nodes by blending dissimilar anchor features, aiming to increase sample diversity and expand the decision boundary of the minority class. Additionally, to prevent the generation of harmful connections, this paper introduces a personalized PageRank method for neighborhood distribution of oversampled samples. Experimental results on three real datasets (Cora, CiteSeer, and Cora-Full) demonstrate a clear advantage of this method in comparison to nine representative approaches. [ABSTRACT FROM AUTHOR]