Reproduce “Graph Neural Networks with Diverse Spectral Filtering” 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
Jingwei Guo, Kaizhu Huang, Xinping Yi, Rui Zhang
The paper introduces Diverse Spectral Filtering (DSF), a framework that enables Graph Neural Networks (GNNs) to learn node-specific spectral filter weights, significantly improving performance on complex, heterogeneous graph datasets. DSF-enhanced models demonstrated up to 4.92% higher accuracy in node classification on heterophilic graphs by integrating iterative positional encoding and local-global weight decomposition.
Key insights
- Enabling node-specific spectral filtering is crucial for effectively handling regional heterogeneity in graphs, leading to improved performance, particularly on heterophilic datasets.
- Position-aware embeddings, generated through an iterative process, can effectively regularize and guide the learning of diverse filter weights, preventing overfitting inherent in directly parameterizing each node's filter.
- Decomposing filter weights into global and local components provides a robust mechanism to capture both general graph properties and fine-grained node-specific characteristics, enhancing model adaptability and interpretability.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.