📣 Our spicy ICML 2025 position paper: “Graph Learning Will Lose Relevance Due To Poor Benchmarks”. Graph learning is less trendy in the ML world than it was in 2020-2022. We believe the problem is in poor benchmarks that hold the field back - and suggest ways to fix it! 🧵1/10

If we're truly making progress, why has the SOTA performance on the Enzymes been declining over the years? 🤔 How is it that replacing the input graph structures with a fixed Cayley graph often performs better? And why are practitioners still relying on vanilla GNNs? 🫣

I've been thinking about the "correct graph" problem for quite some time. There are clearly situations where we expect the graph to encode valuable information for the target variable, yet surprisingly, dropping the graph altogether often yields better performance. More

The best place, Meta TLV 💙, although I am currently at Meta NY 🌝 🗽 Tomorrow I will give a talk at Flatiron Institute, and next week I will be at ICML to present our awesome "Position: Graph Learning Will Lose Relevance Due To Poor Benchmarks" paper! An amazing collab with

I gave a fun talk at Flatiron Institute CCM, met amazing researchers, discussed Foundation Models for graphs, Graph Generative Models, and the future of Graph Learning (which is bright IMHO! 🌞). Many thanks to the best host Teresa Huang for organizing this and arranging me an

Pass by to chat with us on the future of Graph Learning and how we should improve as a research community!!! 😍

So many ppl came to hear the expo on Graph Foundation Models at #ICML2025 by Michael Galkin Bryan Perozzi . This makes me so happy! 🥹 As a contrast to the paper we will present on Thursday, we really should make sure Graph Learning WILL NOT lose relevance😉

After a few conferences of being invited to these events as a candidate, it was so exciting to be one of the hosts of the Meta event at #ICML and tell other AI talented researchers and PhD candidates about our exciting Foundation Models for Graphs research at Meta, and how unique

I tracked the percentage of people who came by our "Position: Graph Learning Will Lose Relevance Due to Poor Benchmarks" poster at ICML 2025 and agreed with the message - 100%! 😄 A recurring question was: But what should we do? Is the field doomed? Definitely not! I see this

😂😂😂 Also, there is just codings, engineering is a relic term from academia. Actually now there is only prompting, coding is a relic term from engineering which is a relic term from academia. 😜