Graph Machine Learning

https://openreview.net/forum?id=SJxzFySKwH

Channel photo updated

Channel name was changed to «Graph Machine Learning»

I had very high expectations for this paper, as I also had some convincing studies showing that you don't need graphs in many datasets. Essentially authors decompose an embedding from the neighborhood into signal and noise and show that with lots of noise you don't need any topology. They also propose an aggregation function that decides on how to filter noise from the neighbors. What's interesting is that they say that mean aggregation is better than sum, while in GIN network the aggregation is proved better with sum instead of mean.

https://openreview.net/forum?id=rkeIIkHKvS

Shows a significant boost in IQ-like tests (originally introduced by deepmind https://deepmind.com/blog/article/measuring-abstract-reasoning) if we use graphs to represent diagrams.

https://openreview.net/forum?id=ByxQB1BKwH

The paper kind of shows that graphs are not necessary for graph classification. If you represent a graph as just a set of nodes without any information on their adjacency and train MLP model, you can get SOTA results. Important lesson to learn when we make judgments about the quality of the idea/paper based on empirical results.

https://openreview.net/forum?id=HygDF6NFPB

My personal favorite from ICLR 2020. The paper shows on which conditions GNN can compute any function and that the product of depth*width of GNN should be of size ~n in order to compute popular statistics on graphs (e.g. diameter, vertex cover, coloring, etc.).

https://openreview.net/forum?id=B1l2bp4YwS