Evaluating Task-Specific Node Influence via Node-Removal-Based Fast Graph Neural Network Inference
Keywords: influential node detection, graph neural network, network analysis
Abstract: Graph neural networks (GNNs) have been widely utilized to capture the underlying information propagation patterns in graph-structured data. While remarkable performance has been achieved in extensive classification tasks, there comes a new trending topic of identifying influential nodes on graphs. This paper investigates a new yet practical problem of evaluating the influence of node existence itself, which aims to efficiently measure the overall changes in the outputs of a trained GNN model caused by removing a node. A realistic example is, ``Under a task of predicting Twitter accounts' polarity, had a particular account not appeared, how might others' polarity be changed?''. A straightforward way to obtain the node influence is to alternately calculate the influence of removing each node, which is reliable but time-consuming. The related lines of work, such as graph adversarial attack and counterfactual explanation, cannot directly satisfy our needs since they typically suffer from low efficiency on large graphs. Besides, they cannot individually evaluate the removal influence of each node. To upgrade the efficiency, we design an efficient algorithm, NOde-Removal-based fAst GNN inference (NORA), which uses the gradient of the neural networks to approximate the node-removal results. It only costs one forward propagation and one backpropagation to approximate the influence score for all nodes. %We also adapt state-of-the-art counterfactual explanation models for our problem.
Extensive experiments are conducted on six benchmark datasets, where {\model} exceeds the compared methods. Our code is available at https://anonymous.4open.science/r/NORA.
Track: COI (submissions co-authored by SAC)
Submission Guidelines Scope: Yes
Submission Guidelines Blind: Yes
Submission Guidelines Format: Yes
Submission Guidelines Limit: Yes
Submission Guidelines Authorship: Yes
Student Author: Yes
Submission Number: 434
Loading