Research On Out Of Distribution Generalized Graph Neural Network Based On Causal Inference

This article mainly uses causal inference methods to study the problem of out-ofdistribution generalization of graph neural networks.In traditional machine learning problems,we generally make the assumption of same distribution,that is,we assume that the training set and test set data follow the same probability distribution.However,this assumption is sometimes impossible to achieve in real-life scenarios,and there is a high possibility that the test set in practical applications is not the same probability distribution as the training set.For example,an image classification model may encounter an image of a category that has never been seen in the training set,and the user of autonomous driving may drive the car into some complex road sections that have never appeared in the training set.This phenomenon where the training set and test set distributions are inconsistent is called out-of-distribution(OOD)phenomenon.Studying out-of-distribution generalization has become a very important issue for future artificial intelligence and deep learning.Out-of-distribution generalization problems are generally studied more in computer vision(CV).The research on out-of-distribution generalization problems in graph neural networks(GNNs)has just begun.In 2021,researchers claimed to have studied the problem of out-of-distribution generalization of graph classification tasks on GNNs for the first time.Starting from the perspective of causal inference and applying the relevant knowledge of causal inference and GNNs,this article proposes a new GNN out-of-distribution generalization model.It compared the proposed model with traditional GNN models on a real dataset,Drug OOD,released by Tencent AI Lab,and demonstrated that our model has superior out-of-distribution generalization ability.This article also demonstrated through ablation experiments that all parts of the model are very important,and the step of extracting causal feature subgraphs improved the accuracy by 17%.This study provides new ideas and methods for the application of graph neural networks in the problem of out of distribution generalization.