Spectral GNNs: Cracking the Illusion of Graph Magic

Spectral Graph Neural Networks (Spectral GNNs) have been touted as the go-to for node classification magic. Their promise? Frequency-domain filtering on graphs. But here's the kicker: their foundations are wobbly at best.

Not Your Classical Fourier

Recent revelations show that the so-called 'graph Fourier bases' aren't true Fourier bases for graph signals. That's a big deal. It means the whole narrative of Spectral GNNs being deeply rooted in graph spectrum processing is as shaky as a house of cards.

And there's more. Those (n-1)-degree polynomials, where n is the number of nodes, can interpolate any spectral response via a Vandermonde system. Translation: the typical 'polynomial approximation' story just doesn't hold water.

Message-Passing Dynamics: The Real MVP

Sources confirm: The success of Graph Convolutional Networks (GCNs) is usually pinned on spectral low-pass filtering. But that's a misconception. The real magic? It's all in the message-passing dynamics, not some mystical Graph Fourier Transform-based spectral wizardry.

Take MagNet and HoloNet. These models were hailed for their spectral prowess. In reality, their effectiveness boils down to implementation quirks that turn them into powerful Message-Passing Neural Networks (MPNNs). Implement them as true spectral algorithms, and the performance tumbles.

Rethinking the Spectral GNN Hype

And just like that, the leaderboard shifts. Spectral GNNs don't truly capture the graph spectrum, nor do they reliably boost performance. Why should anyone care? Because the competitive edge these models claim is often due to their MPNN equivalence, not their spectral intent. It's time to question the hype.

So, why are we still clinging to this spectral fantasy? The labs are scrambling to catch up with this shift in understanding. Maybe it's time the rest of us do the same.