Covid Knowledge Graph

A knowledge graph on COVID-19 that integrates various public datasets. This includes relevant publications, case statistics, genes and functions, molecular data and much more. It's implemented in Neo4j and can be accessed via browser.

The ‘Useless’ Perspective That Transformed Mathematics

Matrix algebra is well understood, while group theory, which is used in many proofs of graph theory and other fields, is much more complicated to study. Representation theory creates a bridge between group theory and linear algebra by assigning a matrix to each element in a group, according to certain rules. This nice article introduces to the world of representation theory.

Fresh picks from ArXiv
This week people share their works they submitted to NeurIPS, a lot of interesting papers from the top people in GML.

GNN
• From Graph Low-Rank Global Attention to 2-FWL Approximation with Yaron Lipman
• Graph Meta Learning via Local Subgraphs with Marinka Zitnik
• Data Augmentation for Graph Neural Networks
• Towards Deeper Graph Neural Networks with Differentiable Group Normalization
• Learning Graph Models for Template-Free Retrosynthesis with Regina Barzilay
• Wide and Deep Graph Neural Networks with Distributed Online Learning
• Pointer Graph Networks with Petar Veličković
• Optimal Transport Graph Neural Networks with Octavian-Eugen Ganea, Regina Barzilay, Tommi Jaakkola
• Manifold structure in graph embeddings
• On the Bottleneck of Graph Neural Networks and its Practical Implications

Conferences
• Exploring Algorithmic Fairness in Robust Graph Covering Problems NeurIPS 2019

Surveys
• A Survey on Generative Adversarial Networks: Variants, Applications, and Training

ICML 2020. Comprehensive analysis of authors, organizations, and countries.

Finally here is my post on the analysis of ICML 2020. There are several things I learned from that. For example that USA participates in 3/4 of the papers 😱 Or that DeepMind makes approximately half of all papers for UK. Or that Google does not collaborate with other companies. Or that, except the USA, there is only China that can brag about several companies that publish regularly. Or that a Japanese professor published 12 papers. And much more.

The code and data is on the github, but the cool part is that you can make your own interactive plots in colab notebook (with no installation required) including a collaboration graph between universities and companies.

Deep learning on graphs: successes, challenges, and next steps

The first blog post of Michael Bronstein about graph learning.

Optimal transport: a hidden gem that empowers today’s machine learning

Very simple explanation of what optimal transport problem is and how it can be applied to various domains such as computer vision. Interestingly just yesterday there was a paper on optimal transport GNN.

June Arxiv: how many graphs papers?

From 18 March to 17 April there were 282 new and 98 updated papers in ArXiv CS section. This is 18 papers less that in the previous period.

PhD Theses on Graph Machine Learning

Here are some PhD dissertations on GML. Part 2 (previous here).

Haggai Marron: Deep and Convex Shape Analysis

Benoit Playe: Machine learning approaches for drug virtual screening

Graph Machine Learning research groups: Tommi Jaakkola

I do a series of posts on the groups in graph research, previous post is here. The eighth is Tommi Jaakkola. He has 7 papers in upcoming ICML 2020. His recent interests include molecular graph design and he maintains AI initiative for finding promising antiviral molecules for COVID-19.


Tommi Jaakkola (~1971)
- Affiliation: MIT
- Education: Ph.D. at MIT in 1997 (supervised by Michael Jordan);
- h-index: 76;
- Awards: Sloan research fellowship, AAAI Fellow;
- Interests: molecular generation, models of GNN

DeepSnap

There is a release of DeepSnap by Stanford group. I have not tested it, but it should allow applying graph algorithms from networkx to pytorch-geometric graphs.

Implicit Neural Representations by Yaron Lipman

The talk Implicit Neural Representations by Yaron Lipman from CVPR 20 workshop on Deep Learning Foundations of Geometric Shape.

Fresh picks from ArXiv
This week highlights applications of GNNs to molecules, contagion, NLP, recommender systems and more.

GNN
• Generalizing Graph Neural Networks Beyond Homophily
• Finding Patient Zero: Learning Contagion Source with Graph Neural Networks with Albert-László Barabási
• MoFlow: An Invertible Flow Model for Generating Molecular Graphs
• Quantifying Challenges in the Application of Graph Representation Learning
• Neural Architecture Optimization with Graph VAE
• Interactive Recommender System via Knowledge Graph-enhanced Reinforcement Learning
• Subgraph Neural Networks with Marinka Zitnik
• Temporal Graph Networks for Deep Learning on Dynamic Graphs with Michael Bronstein
• Erdos Goes Neural: an Unsupervised Learning Framework for Combinatorial Optimization on Graphs with Andreas Loukas
• Walk Message Passing Neural Networks and Second-Order Graph Neural Networks
• Isometric Graph Neural Networks
• Modeling Graph Structure via Relative Position for Better Text Generation from Knowledge Graphs
• Improving Graph Neural Network Expressivity via Subgraph Isomorphism Counting with Michael Bronstein

Math:
• Local limit theorems for subgraph counts
• Longest and shortest cycles in random planar graphs


Conferences
• How to Count Triangles, without Seeing the Whole Graph KDD 2020
• GCC: Graph Contrastive Coding for Graph Neural Network Pre-Training KDD 2020

Surveys
• Localized Spectral Graph Filter Frames: A Unifying Framework, Survey of Design Considerations, and Numerical Comparison

Spektral

Spektral is a library to code GNN in Tensorflow 2 and Keras. New version includes:

- a unified message-passing interface based on gather-scatter
- 7 new GNN layers
- Huge performance improvements
- Improved utils, docs, and examples

The paper will be presented in GRL workshop.

Criteo papers at ICML 2020

Criteo, where I work, this year has record number of accepted papers at ICML. We have 9 papers on various topics, from online learning to theory of optimization to GANs. It makes us 1st company in EU and top-7 company worldwide (among 134 companies who have their papers accepted). So I wrote a short description of each paper in a new blog post.

Top number of submissions at NeurIPS 2020

Mastodons of ML are the following:

* Peter Richtárik (KAUST) 14
* Bernhard Schölkopf (MPI) 13
* Sergey Levine (UC Berkeley) 12
* Masashi Sugiyama (RIKEN) 11
* Yoshua Bengio (MILA) 11

This is based on 2313 arXiv papers that are submitted to NeurIPS2020.

Last year there were at least some people with 15 submissions, so it's probably underestimates these numbers. Also, compared to last year there was 54% of the papers appearing in arXiv at the moment of the conference. For this year, today there are 25% of arXiv papers, so it means not everyone submitted their papers to arXiv.

Sylow theorems and algebraic geometry

There is a fresh thread on Sylow theorems, which are popular results in group theory. I'm not sure how much the waste of time is studying group theory, that's something in my todo list, but this thread is giving a good intro to it.

Bringing Light Into the Dark: A Large-scale Evaluation of Knowledge Graph Embedding Models Under a Unified Framework

This is a post by Michael Galkin (@gimmeblues) about their new work on comprehensive evaluation of knowledge graph embeddings. A lot of interesting insights about knowledge graphs.

Today we are publishing the results of our large-scale benchmarking study of knowledge graph (KG) embedding approaches. Further, we are releasing the code of PyKEEN 1.0 - the library behind the study (in PyTorch)! What makes KGs special: they often have hundreds or thousands of different relations (edge types), and having good representations is essential for reasoning in embedding spaces as well as for numerous NLP tasks.

We often evaluate KG embeddings on the link prediction task - given subject+predicate, the model has to predict most plausible objects. As typical KGs contain 50k-100k different entities, you can guess the top1/top10 ranking task is quite complex!

Why benchmarking is important: currently, there is no baseline numbers to refer to. Lots of papers in the domain are not reproducible, or the authors simply take metrics values as reported in other papers withougt reproducing their results.

In this study, we ran 65K+ experiments and spent 21K+ GPU hours evaluating 19 models spanning from RESCAL first published in 2011 to the late 2019's RotatE and TuckER, 5 loss functions, training strategies with/without negative sampling, and many more hyper-parameters that turn out to be important to consider.

Key findings:
- Careful HPO optimization brings us new SOTA results giving significant gains of 4-5% compared to reported results in respective papers (btw, we used Optuna for HPO);
- Properly tuned classical models (TransE, DistMult) are still good and actually outperform several newer models;
- No Best-of-the-Best Silver Bullet model that beats all others across all tasks - some models better capture transitivity, whereas other better capture symmetric relations;
- Surprisingly, for the inherently ranking task, the ranking loss (or MarginRankingLoss in PyTorch) is suboptimal. Instead, Cross-Entropy and its variations show better result;
- Using all enities for negative sampling, i.e., sigmoid/softmax distribution over all enities, works well but can be quite expensive on large KGs. Stochastic negative sampling is a way to go then;
- Computationally expensive and bigger models do not yield that big and drastic performance gains. In fact, 64-d Rotate is better than most 500-d models.


Paper: https://arxiv.org/abs/2006.13365
Code: https://github.com/pykeen/pykeen