Graph Machine Learning research groups: Christos Faloutsos

I do a series of posts on the groups in graph research. The sixth is Christos Faloutsos. He was an advisor for many current professors in GML such as Jure Leskovec, Leman Akoglu, Stephan Guennemman, and Bruno Ribeiro.


Christos Faloutsos (~1960)
- Affiliation: Carnegie Mellon University; Amazon
- Education: Ph.D. at University of Toronto in 1987 (supervised by Stavros Christodoulakis);
- h-index: 131;
- Awards: ACM fellow, best paper awards at KDD, SIGMOD, ICDM;
- Interests: data mining; database; anomaly detection in graphs.

Social network data set

Anton @xgfsru shared a data set VK1M (password 1234), with first 1M users from social network vk.com (data is taken via public API). In addition to the friends of each user, the file contains meta-information such as education, country, birthday of each node. It can be useful for node classification or regression tasks as well as community or anomaly detection.

Fresh picks from ArXiv
This week has more papers from upcoming ACL, SIGIR, and KDD; a new survey on combinatorial optimization on graphs; and Borsuk's conjecture 🐹

Conferences
• Joint Item Recommendation and Attribute Inference: An Adaptive Graph Convolutional Network Approach SIGIR 20
• ATBRG: Adaptive Target-Behavior Relational Graph Network for Effective Recommendation SIGIR 20
• Connecting the Dots: Multivariate Time Series Forecasting with Graph Neural Networks KDD 20
• Understanding Negative Sampling in Graph Representation Learning KDD 20
• Leveraging Graph to Improve Abstractive Multi-Document Summarization ACL 20
• M2GRL: A Multi-task Multi-view Graph Representation Learning Framework for Web-scale Recommender Systems KDD 20
• Graph Structure Learning for Robust Graph Neural Networks KDD 20


Surveys
• Learning Combinatorial Optimization on Graphs: A Survey with Applications to Networking
• How to Build a Graph-Based Deep Learning Architecture in Traffic Domain: A Survey
• Motif Discovery Algorithms in Static and Temporal Networks: A Survey

Graph Theory
• The Weisfeiler-Leman dimension of distance-hereditary graphs
• Counterexamples to Borsuk's conjecture from a third strongly regular graph
• A Group-Theoretic Framework for Knowledge Graph Embedding

How hard is graph isomorphism for graph neural networks?

This is a new work by Andreas Loukas that sheds a little bit more light into the theory behind GNN. The analysis relies on the amount of information nodes should exchange in order to detect isomorphism class of each graph. This problem of finding isomorphism class is called graph canonization problem, which is probably even harder than graph isomorphism. As such a GNN model needs to output a number for each possible graph up to isomorphism, and needless to say the number of non-isomorphic graphs grows closely to factorial terms. Hence the experiments, while support the theory, are done only on very small graphs ~10 nodes.

On the universal equivariant functions

Yesterday's post was followed by a conversation with Andreas Loukas on the power of graph neural networks. There is one detail about the current analysis of GNN, which I didn't pay much attention before, even though I encountered it (it's a recurring phenomenon for me, when some major insight is given one sentence in the paper and is not highlighted in CAPITAL letters).

The insight is that there are two types of GNN, anonymous and non-anonymous. Anonymous case means you are invariant to the order of nodes, for example when you use the sum aggregation over nodes. It was shown that anonymous case is equivalent to WL algorithm and therefore has a lot limitations such as not being able to count subgraphs or distinguish graphons, etc. So the current anonymous models are not universal: they cannot compute all the functions of the inputs. It's a weaker model than non-anonymous case, when you give some orderings to the nodes and then you iterate over all orderings.

Non-anonymous models have additional node features, for example one-hot encodings of their position in adjacency matrix, which is one of the sufficient conditions for GNN to be universal. There are then two scenarios. Either you consider all possible permutations, in which case it grows in factorial terms and essentially it's a cheat. Or you resort to a single permutation, but then do not enjoy the invariance property of GNN, i.e. for different orderings it can give different set of embeddings for the same nodes.

So it's interesting to see if there are universal equivariant functions that do not use all node permutations, which is still an open question.

Other telegram channels

There are many other interesting channels, below are English-speaking that I follow. If I miss something valuable, feel free to send them to me and I will update the post.

https://t.iss.one/opendatascience
https://t.iss.one/ArtificialIntelligencedl
https://t.iss.one/ai_machinelearning_big_data
https://t.iss.one/j_links
https://t.iss.one/ArtificialIntelligenceArticles
https://t.iss.one/snakers4
https://t.iss.one/loss_function_porn
https://t.iss.one/yegor256news

Also quite useful aggregator https://infomate.club/ml/

Why `True is False is False` -> False?

Stumbled upon this little question why True is False is False evaluates to False in python. The answer is as simple as the question, but not obvious and many people could not answer it. So I wrote a quick post about it.

Fresh picks from ArXiv
This week has GNN variants for various types of graphs, NP-completeness of MaxCut problem, and a survey on graph data management 🖥

GNN
• Understanding the Message Passing in Graph Neural Networks via Power Iteration
• Interpretable and Efficient Heterogeneous Graph Convolutional Network
• Graph Neural Network for Hamiltonian-Based Material Property Prediction
• Non-IID Graph Neural Networks
• Hierarchical Fashion Graph Network for Personalized Outfit Recommendation
• Non-Local Graph Neural Networks

Graph Theory
• Complexity of Maximum Cut on Interval Graphs
• Planar Graphs that Need Four Pages

Surveys
• Benchmarking Graph Data Management and Processing Systems: A Survey

ICML 2020 stats

ICML is the top conference in ML.

Dates: July 12-18
Where: Online

• 4990 submissions (vs 3424 in 2019)
• 1088 accepted (vs 774 in 2019)
• 21.8% acceptance rate (vs 22.6% in 2019)
• 53 graph papers (5% of total)

Of course, there are many big names from graph machine learning for ICML 20 such as Jure Leskovec, Tom Jaakkola, Le Song, Peter Battaglia and others.

Online Seminar on Mathematical Foundations of Data Science

Virtual weekly seminars with excellent list of speakers, open to the public, on mathematics and statistics in ML.

Graph Machine Learning research groups: Joan Bruna

I do a series of posts on the groups in graph research, previous post is here. The seventh is Joan Bruna. He was one of the authors of the survey Geometric Deep Learning: going beyond Euclidean Data and now has increasingly more papers on the theoretical explanations of GNN.


Joan Bruna (~1981)
- Affiliation: New York University
- Education: Ph.D. at Ecole Polytechnique, France in 2013 (supervised by Stephane Mallat);
- h-index: 27;
- Awards: NSF career award, Sloan fellowship, ICMLA best paper;
- Interests: GNN theory, equivariant networks

ICML 2020 arxiv links

There is a nice website that gather all (available) links to papers at ICML. There are some interesting insights.
First, it's interesting to see what is the oldest paper that was accepted to ICML this year. Apparently this paper was published on arxiv in Sep 2017, waiting for a little bit less than 3 years to get accepted. There are 8 papers from 2018. And the authors probably started working on these papers 6-9 months before the publication date. It's brutal.
Another interesting observation is that the word graph appeared to be top-5 word among all words in titles, which show increased interest in graphs at ICML.

Fresh picks from ArXiv
This week is rich on explainable GNN literature, as well as papers on compression of graphs, combinatorial optimization and recommendation.

GNN
• Learning to Solve Combinatorial Optimization Problems on Real-World Graphs in Linear Time
• Are Graph Convolutional Networks Fully Exploiting Graph Structure?
• Accurately Solving Physical Systems with Graph Learning
• Single-Layer Graph Convolutional Networks For Recommendation
• XAI for Graphs: Explaining Graph Neural Network Predictions by Identifying Relevant Walks
• XGNN: Towards Model-Level Explanations of Graph Neural Networks
• Universal Graph Compression: Stochastic Block Models
• Convergence and Stability of Graph Convolutional Networks on Large Random Graphs
• Fairness-Aware Explainable Recommendation over Knowledge Graphs

Graph Theory
• Hierarchical hyperbolicity of graph products
• Tree-Projected Gradient Descent for Estimating Gradient-Sparse Parameters on Graphs
• The Weisfeiler-Leman dimension of chordal bipartite graphs without bipartite claw

Conferences
• Multi-level Graph Convolutional Networks for Cross-platform Anchor Link Prediction KDD 20

Surveys
• Generate FAIR Literature Surveys with Scholarly Knowledge Graphs
• A Comprehensive Survey of Neural Architecture Search: Challenges and Solutions

Gradient Boosting Meets Graph Neural Networks for Heterogeneous Data

We have two short paper submissions this year to GRL workshop this year. One of them is about application of gradient boosting decision trees (GBDT) to graphs. We know that Xgboost, LightGBM, and CatBoost perform extremely well on tabular data and are preferred methods for competitions like Kaggle. But how do you generalize it to graph-structured data?

A naïve approach is to train first GBDT on node features only, ignoring graph topology and then use predictions as additional features to your model. But that misses graph information, possibly leading to inaccurate predictions. Instead, we propose to train GBDT and GNN end-to-end such that each tree of GBDT approximates mistakes made by GNN in the forward passes. We call the model Boosted Graph Neural Network and show that it can lead to significant uplift in performance in node regression task, while being very efficient.

Are Hyperbolic Representations in Graphs Created Equal?

The second submission to GRL workshop was on hyperbolic embeddings for graphs. We first make a good introduction to the distances and dot products in k-Stereographic model (a Riemannian manifold with constant curvature) and fix the issue with taking gradients at zero curvature, by taking a Taylor series expansion around the origin. This allows seamless gradient descent optimization in non-Euclidean space.

Then we make experiments on node and graph classification, link prediction, and graph embedding task (i.e. preserving distances in the latent space) and show that for link prediction and graph embedding there is an uplift in using hyperbolic manifolds, while for node and graph classification Euclidean models work better.

ICML 2020 collaboration graph

As a preview to my future post (next week) about ICML 2020, I want to share a collaboration graph between different organizations. Final graph has 429 nodes (organizations) and 1206 edges (collaborations). Each edge has a weight: the number of papers the organizations collaborated with. As the final graph is too big to display nicely, you can also look at the subgraph between organizations that collaborated the most (at least 30 collaborations). I will release a colab notebook so that you can play with it.