Abstract:One of the key tasks in graph learning is node classification. While Graph neural networks have been used for various applications, their adaptivity to reject option setting is not previously explored. In this paper, we propose NCwR, a novel approach to node classification in Graph Neural Networks (GNNs) with an integrated reject option, which allows the model to abstain from making predictions when uncertainty is high. We propose both cost-based and coverage-based methods for classification with abstention in node classification setting using GNNs. We perform experiments using our method on three standard citation network datasets Cora, Citeseer and Pubmed and compare with relevant baselines. We also model the Legal judgment prediction problem on ILDC dataset as a node classification problem where nodes represent legal cases and edges represent citations. We further interpret the model by analyzing the cases that the model abstains from predicting by visualizing which part of the input features influenced this decision.

What problem does this paper attempt to address?

### What problem does this paper attempt to solve? This paper aims to solve the problem of handling uncertainty in graph neural networks (GNNs) for node classification tasks in high - risk application scenarios. Specifically, the author proposes a new method - **NCwR** (Node Classification with integrated Reject Option), which allows the model to refuse to make predictions when the uncertainty is high. This is especially important in high - risk fields such as legal judgment prediction and medical diagnosis, because wrong predictions may lead to serious consequences. #### Main problems and motivations 1. **Cost of errors in high - risk applications**: - In high - risk applications such as legal judgment prediction, disease prediction, and financial fraud prediction, the cost of wrong predictions is very high. Traditional GNN models are not cost - effective enough when dealing with difficult samples. 2. **Uncertainty estimation**: - There are many uncertainty estimation methods in the literature for high - risk applications, such as confidence - based methods. These methods give a set of labels instead of a single label during prediction by setting confidence parameters to ensure a high probability that the actual label is included in the set of labels. However, these methods have not been fully applied to GNNs. 3. **Reject option classifier**: - The reject option classifier is a method that chooses not to make a decision when facing uncertain or confusing samples. This method has been widely used in fields such as medicine and finance. This paper attempts to introduce the reject option into the node classification task of GNNs. #### Proposed solutions The author proposes two GNN node classification methods based on the reject option: - **NCwR - Cov**: A coverage - based method, with the goal of minimizing the misclassification rate of non - rejected samples while maintaining a certain coverage. - **NCwR - Cost**: A cost - based method, assuming that rejection has a preset cost \(d\), with the goal of learning an optimal node classifier to minimize the total loss. These two methods are respectively applicable to different application scenarios, especially in high - risk fields, and can more flexibly deal with uncertainty and complexity. #### Experimental verification The author conducted experiments on three standard citation network datasets (Cora, Citeseer, Pubmed) and compared them with existing baseline methods. In addition, they also applied this method to the legal judgment prediction task on the Indian Legal Document Corpus (ILDC), further verifying its effectiveness in real - world high - risk scenarios. ### Summary The main contribution of this paper is to introduce the reject option into the node classification task of GNNs, propose a new framework NCwR, which can better handle the uncertainty in high - risk application scenarios. Through experimental verification, this method performs well on multiple datasets, and has important application value especially in high - risk fields such as legal judgment prediction.

Node Classification With Integrated Reject Option

Classifying Nodes in Graphs without GNNs

Node Classification on The Citation Network Using Graph Neural Network

A Novel Two-Channel Classification Approach Using Graph Attention Network with K-Nearest Neighbor

A novel dominating set and centrality based graph convolutional network for node classification

Graph Neural Network with Curriculum Learning for Imbalanced Node Classification

RoCP-GNN: Robust Conformal Prediction for Graph Neural Networks in Node-Classification

A Graph Neural Network-Based Node Classification Model on Class-Imbalanced Graph Data

Graph Neural Networks With Adaptive Confidence Discrimination

Neighborhood Convolutional Network: A New Paradigm of Graph Neural Networks for Node Classification

Graph as a feature: improving node classification with non-neural graph-aware logistic regression

Transfer Learning Under High-Dimensional Graph Convolutional Regression Model for Node Classification

Graph neural networks in node classification: survey and evaluation

GraphAny: A Foundation Model for Node Classification on Any Graph

Active Learning for Graph Neural Networks via Node Feature Propagation.

Bag of Tricks for Node Classification with Graph Neural Networks

Neural Structured Prediction for Inductive Node Classification

CommunityGCN: community detection using node classification with graph convolution network

SelectiveNet: A Deep Neural Network with an Integrated Reject Option

Editable Graph Neural Network for Node Classifications

Effective Semi-Supervised Node Classification on Few-Labeled Graph Data