Graph Unlearning via Reconstruction --- A Range-Null Space Decomposition Approach

Hang Yin, Xiaoyong Peng, Zipeng Liu, Liyao Xiang

Keywords: Graph Unlearning, Range-Null Space Decomposition

TL;DR: efficient graph node unlearning by unlearned embeddings reconstruction

Abstract: Graph unlearning is a machine unlearning technique tailored to graph neural networks (GNNs) to remove nodes or edges from the training graph. Conventional methods such as retraining is highly inefficient, while influence function-based approaches merely work on minor removal, like 10\% or less of the graph edges. To resolve the problems, we reverse the aggregation process in GNN training by modeling the interaction between unlearned nodes and their neighbors. Given one unlearned node, its embedding is roughly disassembled and assigned to its neighbours by reconstruction, and then removed from its neighbours by embedding modification. We also introduce range-null space decomposition to rectify the raw estimation of the interaction with theoretical support. Experimental results on multiple representative datasets and GNN models demonstrate the efficiency of at least $40\times$ acceleration compared with retraining and superior unlearning utility, efficacy, and privacy of our proposed approach compared with other method.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 2405