Graph Unlearning with Efficient Partial Retraining

Graph Neural Networks (GNNs) have achieved remarkable success in various\nreal-world applications. However, GNNs may be trained on undesirable graph\ndata, which can degrade their performance and reliability. To enable trained\nGNNs to efficiently unlearn unwanted data, a desirable solution is\nretraining-based graph unlearning, which partitions the training graph into\nsubgraphs and trains sub-models on them, allowing fast unlearning through\npartial retraining. However, the graph partition process causes information\nloss in the training graph, resulting in the low model utility of sub-GNN\nmodels. In this paper, we propose GraphRevoker, a novel graph unlearning\nframework that better maintains the model utility of unlearnable GNNs.\nSpecifically, we preserve the graph property with graph property-aware sharding\nand effectively aggregate the sub-GNN models for prediction with graph\ncontrastive sub-model aggregation. We conduct extensive experiments to\ndemonstrate the superiority of our proposed approach.\n