topological perception post-processing confidence coefficient correction method applied to GNN

Abstract

The invention discloses a topological perception post-processing confidence coefficient correction method applied to a GNN. The confidence coefficient is corrected by using the homogeneity of the confidence coefficient in the GNN, the correction function is nonlinear transformation, while the classification precision of the original GNN is also saved. Besides, the invention provides a corrected GNN self-training method based on the correction method, the confidence is firstly calibrated by CaGCN and then is used for generating a pseudo tag, high-confidence prediction can be effectively utilized, and the effectiveness of the method provided by the invention in the aspects of correction and accuracy is proved through a large number of experiments.

Description

technical field [0001] The invention relates to the technical field of fiber lasers, in particular to a post-processing confidence correction method applied to GNN topology perception. Background technique [0002] Graphs are ubiquitous in the real world, including social networks, e-commerce networks, transportation networks, etc. Recently, graph neural networks (GNNs) based on message passing have attracted considerable attention in processing graph data. So far, GNNs have been widely used in various applications and achieved remarkable results, such as node classification, link prediction, and graph classification, etc. [0003] In practical applications, however, accuracy is not the only goal. In many applications, especially in the security domain, a trustworthy model is highly desirable. Generally speaking, a trustworthy model means that it should know when its predictions may be incorrect. In other words, the probability (confidence) associated with a predicted cl...

AI Technical Summary

Problems solved by technology

Confidence correction has attracted considerable attention in deep learning, and many studies have shown that modern neural network models are overconfident in predictions, i.e., the prediction accuracy of the model is lower than its confidence

However, the problem of confidence correction has not been studied in GNNs, thus, this raises a fundamental question: Can current GNNs be as overconfident in their predictions as other neural networks?

[0005] In addition, most current self-training methods have the problem of poor performance in high label rate data sets

Considering the low confidence of GCN, we believe that the reason for the poor performance of existing self-training methods is that a large number of high-precision predictions are distributed in the low confidence interval, which prevents them from being added to the training set.

Images