Reproduce “GNNDelete: A General Strategy for Unlearning in Graph Neural Networks” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Jiali Cheng, George Dasoulas, Huan He, Chirag Agarwal, Marinka Zitnik

Researchers from the University of Massachusetts Lowell, Harvard University, and Adobe developed GNNDELETE, a general, model-agnostic strategy for unlearning in Graph Neural Networks that addresses data privacy and model adaptability. It introduces a layer-wise deletion operator and formalizes two key unlearning properties, achieving superior performance on deleted data while being up to 12.3x faster than retraining from scratch and significantly more memory-efficient.

Key insights

  • Effective graph unlearning requires a careful balance between completely erasing the influence of deleted data (Deleted Edge Consistency) and preserving the model's performance on the remaining graph (Neighborhood Influence).
  • A model-agnostic, layer-wise deletion operator applied locally to affected neighborhoods can efficiently unlearn various graph elements without requiring full model retraining or global weight modification.
  • The proposed deletion operator's parameters can be independent of graph size, allowing the strategy to be scalable and computationally efficient for large and evolving graph datasets.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.