A Fast Method for Updating Node Representations Learned by Graph Convolutional Networks When Node Features Change

EasyChair Preprint no. 11131

Abstract

Recently, graph convolutional networks (GCN) have been widely used to analyze graph data in many fields. A GCN takes the features of each node in a given graph and the connections between them as input, and learns a representation for each node. However, one challenge with GCN is that even if only a few nodes' features change over time after training, GCN requires retraining for all nodes. Therefore, in this paper, we propose a method to quickly update the representation of all nodes by utilizing the previous learning results when only some of the nodes in the graph have changed their features. The proposed method quickly computes only the parts that are modified that are influenced by modified features of nodes from the previously learned representation. Experiments using synthetic data show that when the number of nodes with changed features is small, the proposed method significantly reduces the execution time compared to the existing method that retrains all the representations of all nodes.

Keyphrases: Graph Convolutional Network, Graph Neural Network, node representation

@Booklet{EasyChair:11131,
  author = {Yeonju Song and Ki Yong Lee},
  title = {A Fast Method for Updating Node Representations Learned by Graph Convolutional Networks When Node Features Change},
  howpublished = {EasyChair Preprint no. 11131},

  year = {EasyChair, 2023}}