Research On Trustworthy Graph Neural Networks Based On The Confidence

With the in-depth development of deep learning technology,the Graph Neural Networks(GNNs),a type of representation learning method for for graphs,has received extensive attention.Experiments show that GNNs have achieved excellent results in a variety of graph-related tasks and scenarios,ranging from social network analysis,traffic flow prediction to biomedicine,recommendation systems and computer vision.Although the GNNs have exhibited its superiority,whether their predictions are trustworthy remains to be explored,especially in the risk-sensitive scenarios,where the trustworhtiness becomes more urgent.So far,there is still no consensus on the definition of trustworthiness,and various research field,such as the robustness,interpretability,and out-of-distribution generalization,are all considered as one of the aspects of trustworthiness.Among them,there are some studies that believe a trustworthy model should have self-awareness of its predictive ability by virtue of the trustworthy confidence.In other words,a trustworthy model should give high confidence to its correct predictions and low confidence to its incorrect predictions,that is,the confidence should be consistent with the accuracy of predictions.Previous studies have shown that many neural networks are over-confident in their predictions,that is,the average confidence in the predictions are generally higher than the average prediction accuracy.It implies the confidence of existing neural networks are not trustworthy.Based on this,this study aims to explore the trustworthiness of the GNNs from the perspective of confidence.This paper first explores whether the confidence of the GNNs are consistent with their prediction accuracy.The investigation indicates that existing GNNs are far distant from being trustworthy,and surprisingly show that the confidence of GNNs in their predictions are generally lower than their average prediction accuracy,that is,GNNs tend to be underconfident.Based on this,this paper designs a topology-aware post-hoc confidence calibration function to improve the trustworthiness of confidence.The calibration function utilizes the well-known graph convolutional neural network as the backbone model,and incorporates the temperature scaling method to ensure that the predictive performance of the GNNs will not be changed after confidence calibration,leading to a novel trustworthy GNN model,CaGCN.Finally,this paper demonstrates the effectiveness of CaGCN in improving the trustworthiness of confidence through extensive experiments.Inspired by the discovery above that GNNs are untrustworthy,this paper further explores the trustworthiness of confidence in graph selftraining learning,an important application for confidence.Results show that introducing redundant high-confidence predictions during graph selftraining could not bring more gain.Worse still,it could introduce the distribution shift for the training set.To sum up,confidence in the graph self-training learning is untrustworthy.Based on this,this paper proposes a novel graph self-training framework DR-GST,where the proposed information gain-based loss function can appropriately address the distribution shift,and the additional loss correction strategy could improve pseudo-labels.Finally,this paper demonstrates the rationality and effectiveness of DR-GST through a theoretical analysis and extensive experiments.In addition,note that the trustworthy GNN could still obtain untrustworthy predictions because of the noise and privacy protection during collecting graph data in reality.Therefore,this study proceeds to explore how to improve the trustworthiness of graph data,i.e.,the input of GNNs,from the perspective of confidence.Considering the decisive role of the message passing mechanism on GNNs,this papers starts with the node degree,which is the key factor for the message passing mechanism,and explores how to obtain trustworthy topology for a long-tailed graph.In this line,this paper proposes a confidence-guided trustworthy topology generator,which can generate local graph topology for tail nodes with the least information under the guidance of the informative head nodes.The paper concludes by demonstrating the effectiveness of the proposed TNTG in improving GNNs on node classification,especially tail node classification.