Graph Convolutions That Ditch the Eigendecomposition

What it does

This is a PyTorch reimplementation of the Graph Wavelet Neural Network (GWNN), a spectral graph convolutional network built for semi-supervised node classification. Instead of using the graph Fourier transform—which requires an expensive matrix eigendecomposition—it performs convolutions in the wavelet domain. The authors report that this approach outperforms earlier spectral graph CNNs on standard citation benchmarks like Cora, Citeseer, and Pubmed.

The interesting bit

Graph wavelets are sparse and localized in the vertex domain, so you get interpretable, efficient convolutions without diagonalizing the Laplacian. The code leans on PyGSP for the wavelet machinery and torch-geometric for the graph backend.

Key highlights

• Replaces graph Fourier transform with graph wavelet transform to avoid matrix eigendecomposition.
• Targets semi-supervised node classification on citation graphs (Cora, Citeseer, Pubmed).
• Sparse binary feature support via JSON and edge-list CSV inputs.
• Exposes hyperparameters like Chebyshev polynomial order and heat kernel scale for tuning the wavelet approximation.
• Includes a link to a reference TensorFlow implementation for cross-checking.

Caveats

• The README pins very old dependency versions (Python 3.5.2, PyTorch 1.1.0), so expect friction with modern toolchains.
• It is a straight research reimplementation; the repo does not claim novel extensions or production hardening.

Verdict

Worth a look if you are studying spectral GNNs or need a working baseline for node classification. Skip it if you want a maintained, plug-and-play graph learning library.