Graph Machine Learning
6.71K subscribers
53 photos
11 files
808 links
Everything about graph theory, computer science, machine learning, etc.


If you have something worth sharing with the community, reach out @gimmeblues, @chaitjo.

Admins: Sergey Ivanov; Michael Galkin; Chaitanya K. Joshi
Download Telegram
Graph Representation Learning for Drug Discovery Slides

Slides from Jian Tang of the talk on de novo drug discovery and drug repurposing.
Invariant and equivariant layers with applications to GNN, PointNet and Transformers

A blog post by Marc Lelarge about invariant and equivariant functions and their relation to the universality and expressivity of GNN. As the main result they show that any invariant/equivariant function on n points can be represented as a sum of functions on each point independently.
Video: Workshop of Graph Neural Networks and Systems (GNNSys'21)

Very interesting videos from the workshop at MLSys 21 on GNNs in the industry. The talks include topics such as GNNs on graphcore's IPU, chip placement optimization, particle reconstruction at the large hadron collider and more.
GML In-Depth: three forms of self-supervised learning

My new in-depth newsletter on self-supervised learning with applications to graphs. There is an upcoming keynote talk from Alexei Efros at ICLR'21 about self-supervised learning and I was inspired by the motivations that he talks there. In particular, he explains that self-supervised learning is a way to reduce the role of humans in designing ML pipelines, which would allow neural nets to learn in a similar way as humans do. Self-supervised learning for graphs is an active area of research and there are good reasons for this: for applications such as drug or catalyst discovery, there are billions of unlabeled graphs from which we would like to extract as much relevant information as possible. So self-supervised learning is becoming a new paradigm for learning such useful representations.
Knowledge Graphs @ ICLR 2021

One and only Michael Galkin does it again with a superior digest of knowledge graph research at ICLR 2021. Topics include reasoning, temporal logics, and complex question answering in KGs: a lot of novel ideas and less SOTA-chasing work!
Fresh picks from ArXiv
This week on ArXiv: optimization properties of GNNs, review on sample-based approaches, and time zigzags for Ethereum price prediction ๐Ÿ’ฐ

If I forgot to mention your paper, please shoot me a message and I will update the post.

Conferences
* Unsupervised Cross-Domain Prerequisite Chain Learning using Variational Graph Autoencoders ACL 2021
* Neural Graph Matching based Collaborative Filtering SIGIR 2021
* Z-GCNETs: Time Zigzags at Graph Convolutional Networks for Time Series Forecasting ICML 2021
* Optimization of Graph Neural Networks: Implicit Acceleration by Skip Connections and More Depth ICML 2021

Efficiency
* Scalable Graph Neural Network Training: The Case for Sampling
* VersaGNN: a Versatile accelerator for Graph neural networks
New Proof Reveals That Graphs With No Pentagons Are Fundamentally Different

A new article at Quanta about Erdล‘sโ€“Hajnal conjecture, which states that any graph that forbids having some small subgraph will inevitably have a large clique or a large independent set. The article talks about a recent paper that confirms the conjecture for a special case which was deemed the hardest. Now there is a hope that the conjecture is true for the general case.
Constructions in combinatorics via neural networks

I have been fascinated about potential of using machine learning for combinatorial problems and have written multiple posts (here and here) and a survey about this. And as such it was exciting to see a work that applies RL framework to disprove several combinatorial conjectures.

The algorithm is very simple: generate many graphs with MLP, select the top-X of them, use cross-entropy to update MLP. So it does not use recent advances in RL, neither in GML to care about invariance of the input. So there is a room for improvement. Also it generates graphs of pre-determined size, so if a counterexample has a big order it would be difficult to know in advance. But it would be very interesting to apply this framework to more complicated conjectures such as reconstruction conjecture.
On Explainability of Graph Neural Networks via Subgraph Explorations

This is a guest post by Shuiwang Ji about their recent work, accepted to ICML 2021.

Title: "On Explainability of Graph Neural Networks via Subgraph Explorations"

TL; DR:
- We propose a novel method, known as SubgraphX, to explain GNNs by exploring and identifying important subgraphs.
- We propose to incorporate the Monte Carlo tree search to explore subgraphs and propose efficient approximation schemes to measure subgraphs via Shapley values.
- Our proposed method consistently and significantly outperforms state-of-the-art techniques.
Code is now available as part of our DIG library.

We study the explainability of Graph Neural Networks and propose a novel method (SubgraphX) to provide subgraph-level explanations. While existing methods mainly focus on explaining GNNs with graph nodes or edges, we argue that subgraphs are more intuitive and human-intelligible.

In our SubgraphX, we propose to explore different subgraphs with the Monte Carlo tree search. For each subgraph, we measure its importance using Shapley values, which can capture the interactions among different graph structures. We further improve the efficiency with our proposed approximation schemes to compute Shapley values for graph data. Both quantitative and qualitative studies show our method obtain higher-quality and more human-intelligible explanations while keeping time complexity acceptable.

Our method represents the first attempt to explain GNNs by explicitly studying the subgraphs. We hope that this work can provide a new direction for the community to investigate the explainability of GNNs in the future.
GraphDF: A Discrete Flow Model for Molecular Graph Generation

This is a guest post by Shuiwang Ji about their recent work, accepted to ICML 2021.

Title: โ€œGraphDF: A Discrete Flow Model for Molecular Graph Generationโ€

TL; DR:
- We propose GraphDF, a novel discrete latent variable model for molecular graph generation method.
- We propose to use invertible modulo shift transform to sequentially generate graph nodes and edges from discrete latent variables.
- Our proposed method outperforms prior methods on random generation, property optimization, and constrained optimization tasks.
Code is now available as part of our DIG library.

We study the molecular generation problem and propose a novel method (GraphDF) achieving new state-of-the-art performance. While prior methods use continuous latent variables, we argue that discrete latent variables are more suitable to model the categorical distribution of graph nodes and edges.

In our GraphDF, the molecular graph is generated by sequentially using modulo shift transform to convert a sampled discrete latent variable to the categorical number of the graph node or edge type. The use of discrete latent variables eliminates the bad effect of dequantization and models the underlying distribution of graph structures more accurately. The modulo shift transform captures conditional information from the last sub-graph by graph convolutional networks to ensure the order invariance. Comprehensive studies show that our method outperform prior methods on random generation, property optimization, and constrained optimization tasks.

Our method is the first work to model the density of complicated molecular graph data with discrete latent variables. We hope that it can provide a new insight for the community to explore more powerful graph generation models in the future.
Rethinking Graph Neural Architecture Search from Message-passing

With abundance of GNNs architectures it's natural to ask how to select the right architecture for your task. In a recent CVPR 2021 work propose a generic architecture that encompasses many existing GNNs, which is then optimized via gradient descent. After optimization resulted GNNs may get different architectures for each layer of GNNs.
Graph Machine Learning research groups: Yizhou Sun

I do a series of posts on the groups in graph research, previous post is here. The 28th is Yizhou Sun, a professor at UCLA, who co-authored a book on heterogeneous information networks.

Yizhou Sun (~1982)
- Affiliation: UCLA
- Education: Ph.D. at UIUC in 2012 (advisors: Jiawei Han)
- h-index 48
- Interests: heterogeneous information networks, self-supervised learning, community detection
- Awards: best research papers at KDD, ASONAM
Mathematicians Answer Old Question About Odd Graphs

A new post at Quanta about the work that settles the question (c. 1960s) of the biggest subgraph with all vertices having odd degree within that subgraph.
NAACL-2021 Papers

A list of accepted papers to NLP conference NAACL-2021 is available at digest console. There are ~40 graph papers out of 476 papers.