antonisva

Spectral and message-passing GNNs are often studied by different communities, with little interaction. In arxiv.org/abs/2602.10031, we argue that the two architecture types have distinct strengths and should be studied in a unified way.

Recently, we have been looking into putting neural algorithmic reasoning on theoretical footings. (1) In arxiv.org/abs/2602.13106, we derive sufficient conditions for GNNs to size generalize algorithms from finite data.

Fun times at ICML. Graph learning dinner, position poster gang, theory, and graph learning hike. :)

At ICML 🇨🇦 presenting the spicy 🌶️ Position: Graph Learning Will Lose Relevance Due To Poor Benchmarks 📍 East Hall A-B #E-604, Thu Also, @antvas98 will be presenting "Covered Forest" — glad to have played a part in this one! 📍 #E-2908, Thu DM to chat graph(+foundation models)

Want to know about the current understanding of the generalization abilities of GNNs? Please have a look at our survey paper arxiv.org/abs/2503.15650. Joint work with Antonis Vasileiou, Stefanie Jegelka, and @levie_ron.

In arxiv.org/abs/2412.07106, we leverage modern graph similarity—capturing the fine-grained geometry of MPNNs' feature space—to derive generalization bounds for MPNNs. Our theory is broad, covering various aggregation and loss functions.