Reproduce “Label Deconvolution for Node Representation Learning on Large-scale Attributed Graphs against Learning Bias” 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
Zhihao Shi, Jie Wang, Fanghua Lu, Hanzhu Chen, Defu Lian, Zheng Wang, Jieping Ye, Feng Wu
Label Deconvolution (LD) introduces an efficient label regularization technique to mitigate learning bias when separately training node encoders and Graph Neural Networks on large-scale attributed graphs. This method approximates an inverse GNN mapping to generate structure-aware 'inverse labels' for node encoder supervision, achieving superior prediction accuracy and efficiency on OGB benchmarks compared to existing methods.
Key insights
- Approximating the inverse mapping of GNNs allows for the creation of 'inverse labels' that provide node encoders with supervision signals implicitly pre-adjusted for graph structure, thereby alleviating learning bias.
- The GNN's linear spectral filter can be effectively inverted by expressing it as a polynomial of the normalized adjacency matrix, enabling a tractable and trainable label deconvolution process.
- Blending true node labels with structure-aware inverse labels during NE training offers a 'de-noised' supervision signal, ensuring learned features are optimally compatible with subsequent GNN processing for improved performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.