1

CS6208 : Advanced Topics in Artificial Intelligence

Graph Machine Learning

Lecture 1 : Introduction to Graph Machine Learning

Semester 2 2022/23

Xavier Bresson
https://twitter.com/xbresson

Department of Computer Science

National University of Singapore (NUS)

Xavier Bresson 1
 2

Outline

The Deep Learning (DL) revolution

Limitations of DL
Graph-Structured data
Graph Neural Networks (GNNs)
GNN case studies
GNN for industry
GNN libraries
Conclusion

Xavier Bresson 2
 3

Outline

The Deep Learning (DL) revolution

Limitations of DL
Graph-Structured data
Graph Neural Networks (GNNs)
GNN case studies
GNN for industry
GNN libraries
Conclusion

Xavier Bresson 3
 4

AI : Birth, winter and renaissance

Hardware
GPU speed 2x/1.5 year
CNN
LeCun-Bengio Big Data
Volume 2x/1.5 year
First RNN/LSTM
GAN, Goodfellow
NeurIPS Schmidhuber
ResNet, He
Deep RL, Mnih Google AI
Visual primary cortex First Attention, Bengio TensorFlow
Hubel-Wiesel KDD First Amazon Kaggle Facebook AI
1959 1989 1997 1998 Cloud Center Platform PyTorch

1958 1962 1986 1987 1995 1999 2006 2010 2012 2014 2015
Perceptron Backprop Facebook Center Data scientist
Rosenblatt Hinton-Rumelhart OpenAI Center 1st Job in US

Neocognitron
Fukushima
Birth of First NVIDIA 4th industrial revolution?
Data Science GPU Deep Learning Digital intelligence
Split from Statistics Breakthrough revolution
Tukey Hinton-Krizhevsky or new AI bubble?
SVM/Kernel techniques
Vapnik

AI Birth AI Winter [1960’s-2012] AI Renaissance

Kernel techniques
Graphical models
Wavelets, numerical PDEs
Xavier Bresson Handcrafted Features 4
 5

Supervised big data

Google’s Law of Data : Volume of data double every

1.5 years.
Eric Schmidt : “Every 2 days we create as much
information as we did up to 2003.”

Data generation/collection is easy but data labeling

is challenging.
It is time and money consuming.
It requires humans (machine ability bounded by Performance on
human intelligence). your specific task

Huge neural network

Why do we need labeled/annotated data? Because Big neural network

supervised deep learning is designed to predict labels. Medium neural network
Traditional algorithms

(Big) data

(Labeled) data available for your specific task

Xavier Bresson 5
 6

Self-supervised big data

How much labeled data to achieve super-human

performances?
ImageNet for image recognition has 1.4M images
(machine error is 3% and human error is 5%).
AlphaGo Zero played 4.9 million games.
Self-driving cars : Tesla’s self-driving cars collected 2
billion miles in 2019 (and no autonomous cars so far).

The next AI revolution will not be supervised !

It will be self-supervised with e.g. LLMs/ChatGPT.

Xavier Bresson 6
 7

Computational power

1975 Moore’s Law of Computers :

100x faster every decade
CPU, GPU/TPU, cloud computing
Brain : 1014 synapses/103 Tflops = (standard) computers
in 2030 (100 times faster than today)

1st Amazon Cloud in 2006

1st Intel CPU 1971 2030

=
1st Nvidia GPU in 1999

Xavier Bresson 7
 8

Neural network architectures

Most research works focus on deep learning architectures (along with training and optimization)
Industry labs leverage academic ideas to large-scale/impressive experiments (e.g. generative
models)
Still, no architecture can solve simultaneously many tasks like the human brain but progress has
been made e.g. Transformers.

Biological Neural Network

Artificial Neural Networks

Xavier Bresson 8
 9

The deep learning revolution

Breakthrough in Computer Vision (CNNs)
LeCun, Bottou, Bengio, Haffner 1998
Krizhevsky, Sutskever, Hinton, 2012

Human accuracy 4.1%

Feature Engineering (SIFT) Network engineering (End-to-end sys)

Breakthrough in Speech and Natural Language Processing

(RNNs/Transformers)
Hochreiter, Schmidhuber, 1997 and Vaswani et-al, 2017
Xavier Bresson 9
 10

The power of deep learning

CNNs/RNNs/Transformers are powerful
architectures to solve high-dimensional learning
problems.

Curse of dimensionality :
dim(image) = dim(1000 x 1000) = 106
For N=10 samples/dim ⇒ 101,000,000 points

Xavier Bresson 10
 11

Applications

Computer Vision : Autonomous driving, Face recognition

Natural Language Processing : Machine Translation, Text generation, Chatbot (ChatGPT)
Speech Recognition : Virtual assistants (Alexa/Siri/Google/Cortana)

Xavier Bresson 11
 12

Outline

The Deep Learning (DL) revolution

Limitations of DL
Graph-Structured data
Graph Neural Networks (GNNs)
GNN case studies
GNN for industry
GNN libraries
Conclusion

Xavier Bresson 12
 13

Limitations of deep learning

2 3
r
4 g 5 2 R3
b
<latexit sha1_base64="w4QopqMsjQUq95M2uAzpYQsjt/o=">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</latexit>

Standard DL such as:

CNNs, RNNs, Transformers
grid
Modern DL toolboxes (TensorFlow, PyTorch) Computer Vision (CV)
Successful products for CV/NLP/SR (Google, Meta, etc.)
but limited to work with grid/sequence data. ⇥
<latexit sha1_base64="RbBCGoFPOhMaEQWjMGIngkOT+zo=">AAACJnicbVBNS8NAEN34WeNX1aOXYBE8lUQEvQgFLx4VbCs2oWw2k3Zxswm7E6WE/Bov/hUvHhQRb/4UN20Ofg0MPN6b4c28MBNco+t+WHPzC4tLy40Ve3VtfWOzubXd02muGHRZKlJ1HVINgkvoIkcB15kCmoQC+uHtWaX370BpnsornGQQJHQkecwZRUMNm6e+gBgHth/CiMuCKkUnZSFEad/bPsioZmxf8dEYA9vn0k8ojsOwuCntYbPltt1pOX+BV4MWqeti2Hzxo5TlCUhkgmo98NwMA2OCnAkwNrmGjLJbOoKBgZImoINi+mbp7BsmcuJUmZboTNnvGwVNtJ4koZmsTtS/tYr8TxvkGJ8EBZdZjiDZzCjOhYOpU2XmRFwBQzExgDLFza0OG1NFGZpkqxC83y//Bb3Dtue2vcujVuegjqNBdskeOSAeOSYdck4uSJcw8kCeyAt5tR6tZ+vNep+Nzln1zg75UdbnFzpJpiA=</latexit>
w
⇤
2Z

sequence
Do we have data NOT based on grid/sequence?
Yes, quite a lot! (see next slide)
Natural Language Processing (NLP)

How to generalize NNs beyond CV/NLP/SR? ⇥

<latexit sha1_base64="13AHF1EgpZ6srhwgByuzvfsJlwo=">AAACJXicbVBNS8NAEN34WeNX1aOXYBF6KokIevBQ8OKxiv2AJpTNdtIu3WzC7kQooX/Gi3/FiweLCJ78K27bHLR1YODx3gxv5oWp4Bpd98taW9/Y3Nou7di7e/sHh+Wj45ZOMsWgyRKRqE5INQguoYkcBXRSBTQOBbTD0e1Mbz+B0jyRjzhOIYjpQPKIM4qG6pVvfAERdm0/hAGXOVWKjie5EBM7tX2Q/YKxfcUHQwxsn0s/pjgMw/xh0itX3Jo7L2cVeAWokKIavfLU7ycsi0EiE1TrruemGBgP5EyAcck0pJSN6AC6Bkoagw7y+ZcT59wwfSdKlGmJzpz9vZHTWOtxHJrJ2YV6WZuR/2ndDKPrIOcyzRAkWxhFmXAwcWaROX2ugKEYG0CZ4uZWhw2pogxNsLYJwVt+eRW0LmqeW/PuLyv1ahFHiZySM1IlHrkidXJHGqRJGHkmr+SdTK0X6836sD4Xo2tWsXNC/pT1/QPWxKX9</latexit>
p
⇤
2R

sequence
How to design universal and broadly applicable architectures?
Speech Recognition (SR)
A solution is Graph Neural Networks (GNNs)
With universal architecture (unified all NNs) – New NN generation

Xavier Bresson 13
 14

Outline

The Deep Learning (DL) revolution

Limitations of DL
Graph-Structured data
Graph Neural Networks (GNNs)
GNN case studies
GNN for industry
GNN libraries
Conclusion

Xavier Bresson 14
 2 3 15
type
Graph-structured data i
<latexit sha1_base64="RlAIlE5MorzaIuoAsYZUvwE4eUI=">AAAB6nicbVBNS8NAEJ3Urxq/qh69LBbBg5REBD0WvHisaD+gDWWznbRLN5uwuxFK6U/w4kERr/4ib/4bN20O2vpg4PHeDDPzwlRwbTzv2ymtrW9sbpW33Z3dvf2DyuFRSyeZYthkiUhUJ6QaBZfYNNwI7KQKaRwKbIfj29xvP6HSPJGPZpJiENOh5BFn1Fjpgbtuv1L1at4cZJX4BalCgUa/8tUbJCyLURomqNZd30tNMKXKcCZw5vYyjSllYzrErqWSxqiD6fzUGTmzyoBEibIlDZmrvyemNNZ6Eoe2M6ZmpJe9XPzP62YmugmmXKaZQckWi6JMEJOQ/G8y4AqZERNLKFPc3krYiCrKjE0nD8FffnmVtC5rvlfz76+q9YsijjKcwCmcgw/XUIc7aEATGAzhGV7hzRHOi/PufCxaS04xcwx/4Hz+ADHkjP8=</latexit>
<latexit
Atomi 4 coordinates 5 2 Rdv
charge i <latexit sha1_base64="eM9FjXFYt9MdTuvIcqbvfphanyk=">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</latexit>

ij
<latexit sha1_base64="f8VGxI+5fMbgb8XS6gM2gFgsuvQ=">AAAB63icbVBNS8NAEJ3Urxq/qh69LBbBg5REBD0WvHisYD+gDWWz3bRrdzdhdyKU0r/gxYMiXv1D3vw3Jm0O2vpg4PHeDDPzwkQKi5737ZTW1jc2t8rb7s7u3v5B5fCoZePUMN5ksYxNJ6SWS6F5EwVK3kkMpyqUvB2Ob3O//cSNFbF+wEnCA0WHWkSCUcwl8ei6/UrVq3lzkFXiF6QKBRr9yldvELNUcY1MUmu7vpdgMKUGBZN85vZSyxPKxnTIuxnVVHEbTOe3zshZpgxIFJusNJK5+ntiSpW1ExVmnYriyC57ufif100xugmmQicpcs0Wi6JUEoxJ/jgZCMMZyklGKDMiu5WwETWUYRZPHoK//PIqaV3WfK/m319V6xdFHGU4gVM4Bx+uoQ530IAmMBjBM7zCm6OcF+fd+Vi0lpxi5hj+wPn8AfixjXM=</latexit>

⇢
1 if ij bond
2 3 j Bondij Aij = 0 otherwise
messages
<latexit sha1_base64="IvN7ZIjsPg5Gkwiu2YyISSTnuhw=">AAAB6nicbVBNS8NAEJ34WetX1aOXxSJ4kJKIoMeCF48V7Qe0oWy2m3btZhN2J0IJ/QlePCji1V/kzX/jps1BWx8MPN6bYWZekEhh0HW/nZXVtfWNzdJWeXtnd2+/cnDYMnGqGW+yWMa6E1DDpVC8iQIl7ySa0yiQvB2Mb3K//cS1EbF6wEnC/YgOlQgFo2il+8dyuV+pujV3BrJMvIJUoUCjX/nqDWKWRlwhk9SYrucm6GdUo2CST8u91PCEsjEd8q6likbc+Nns1Ck5tcqAhLG2pZDM1N8TGY2MmUSB7Ywojsyil4v/ed0Uw2s/EypJkSs2XxSmkmBM8r/JQGjOUE4soUwLeythI6opQ5tOHoK3+PIyaV3UPLfm3V1W6+dFHCU4hhM4Aw+uoA630IAmMBjCM7zCmyOdF+fd+Zi3rjjFzBH8gfP5AzNqjQA=</latexit>

<latexit sha1_base64="bTNTV9FCbrI+njHorojZrqu7rQ4=">AAACU3icbVFNa9wwFJTdJE2Vj27bYy6iS0JOix0C7aWQkkuPKWSTwGpZZPl5VxtZMtJz28X4P4ZCD/0jvfTQyhsf0qQDgmHmfUijrNLKY5L8jOJnG5tbz7df0J3dvf2Xg1evr7ytnYSxtNq6m0x40MrAGBVquKkciDLTcJ3dnnf+9RdwXllziasKpqWYG1UoKTBIs8GSfpw1atl+4BoK5A3lGcyVaYRzYtU2WsuWpuyII3xDVzZMFUwtWWZN3jLOafLAsrgA91V5aCkHk/cjKHdqvsARpbPBMBkla7CnJO3JkPS4mA2+89zKugSDUgvvJ2lS4TTMRSV1t6X2UAl5K+YwCdSIEvy0WWfSssOg5KywLhyDbK0+7GhE6f2qzEJlKXDhH3ud+D9vUmPxftooU9UIRt4vKmrN0LIuYJYrBxL1KhAhnQp3ZXIhnJAYvqELIX385Kfk6mSUJqP08+nw7LiPY5sckLfkmKTkHTkjn8gFGRNJ7sgv8ici0Y/odxzHG/elcdT3vCH/IN77C0lLsuA=</latexit>

4 images 5 2 Rd 
Useri type
videos 2 Rd e
i <latexit sha1_base64="1/Uep78HsaIwPrl6syD3IdwJCsQ=">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</latexit>
energy ij
<latexit sha1_base64="j1AqJOvHFkVK1adBuFYlSGbt+wQ=">AAACVHicbVHBTtwwEHUCtDS0dEuPvUSskDitkqpSe1yplx6hYgFpHSLHmWTddZzInlSNrHxkOSDxJb1wwFmiqgWeZOnNm3me8ThrpDAYRbeev7W98+Ll7qtg7/Wb/beTdwfnpm41hwWvZa0vM2ZACgULFCjhstHAqkzCRbb+OuQvfoI2olZn2DWQVKxUohCcoZPSyZpKKHAZ0AxKoSzTmnW95bwPKMIv1JUdXD2lf2NQoMtuo4DKR0dAtShXmKRW/HCBULRiuMoy+72/snkKfeCQTqbRLNogfErikUzJiJN0ck3zmrcVKOSSGbOMowYT1xIFl+5O2hpoGF+zEpaOKlaBSexmKX145JQ8LGrtjsJwo/7rsKwypqsyVznMah7nBvG53LLF4ktihWpaBMUfGhWtDLEOh12FudDAUXaOMK6FmzXkK6YZR/cPwxLix09+Ss4/zuJoFp9+ms6Px3Xskg/kkByTmHwmc/KNnJAF4eQ3+eMRz/NuvDt/y995KPW90fOe/Ad//x6Y+LTc</latexit>

⇢ 2 3
User 1 if ij friends type
i Aij = Atomj 4 coordinates 5 2 Rdv
connectionij 0 otherwise Quantum Chemistry
<latexit sha1_base64="RlAIlE5MorzaIuoAsYZUvwE4eUI=">AAAB6nicbVBNS8NAEJ3Urxq/qh69LBbBg5REBD0WvHisaD+gDWWznbRLN5uwuxFK6U/w4kERr/4ib/4bN20O2vpg4PHeDDPzwlRwbTzv2ymtrW9sbpW33Z3dvf2DyuFRSyeZYthkiUhUJ6QaBZfYNNwI7KQKaRwKbIfj29xvP6HSPJGPZpJiENOh5BFn1Fjpgbtuv1L1at4cZJX4BalCgUa/8tUbJCyLURomqNZd30tNMKXKcCZw5vYyjSllYzrErqWSxqiD6fzUGTmzyoBEibIlDZmrvyemNNZ6Eoe2M6ZmpJe9XPzP62YmugmmXKaZQckWi6JMEJOQ/G8y4AqZERNLKFPc3krYiCrKjE0nD8FffnmVtC5rvlfz76+q9YsijjKcwCmcgw/XUIc7aEATGAzhGV7hzRHOi/PufCxaS04xcwx/4Hz+ADHkjP8=</latexit>
<latexit

charge
<latexit sha1_base64="YFFfb+0iRyjOOtARB3DU8MK64Q0=">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</latexit>
sha1_base64="hP+6LrUf2d3tZaldqaQQvEKMXyw=">AAAB2XicbZDNSgMxFIXv1L86Vq1rN8EiuCozbnQpuHFZwbZCO5RM5k4bmskMyR2hDH0BF25EfC93vo3pz0JbDwQ+zknIvSculLQUBN9ebWd3b/+gfugfNfzjk9Nmo2fz0gjsilzl5jnmFpXU2CVJCp8LgzyLFfbj6f0i77+gsTLXTzQrMMr4WMtUCk7O6oyaraAdLMW2IVxDC9YaNb+GSS7KDDUJxa0dhEFBUcUNSaFw7g9LiwUXUz7GgUPNM7RRtRxzzi6dk7A0N+5oYkv394uKZ9bOstjdzDhN7Ga2MP/LBiWlt1EldVESarH6KC0Vo5wtdmaJNChIzRxwYaSblYkJN1yQa8Z3HYSbG29D77odBu3wMYA6nMMFXEEIN3AHD9CBLghI4BXevYn35n2suqp569LO4I+8zx84xIo4</latexit>
sha1_base64="XmNjvNjNjCnvMqRqnqHTYrXDUkU=">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</latexit>
sha1_base64="VhfVKpfCiyy+XDZ05WLJJbP4cr8=">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</latexit>

ij
<latexit sha1_base64="f8VGxI+5fMbgb8XS6gM2gFgsuvQ=">AAAB63icbVBNS8NAEJ3Urxq/qh69LBbBg5REBD0WvHisYD+gDWWz3bRrdzdhdyKU0r/gxYMiXv1D3vw3Jm0O2vpg4PHeDDPzwkQKi5737ZTW1jc2t8rb7s7u3v5B5fCoZePUMN5ksYxNJ6SWS6F5EwVK3kkMpyqUvB2Ob3O//cSNFbF+wEnCA0WHWkSCUcwl8ei6/UrVq3lzkFXiF6QKBRr9yldvELNUcY1MUmu7vpdgMKUGBZN85vZSyxPKxnTIuxnVVHEbTOe3zshZpgxIFJusNJK5+ntiSpW1ExVmnYriyC57ufif100xugmmQicpcs0Wi6JUEoxJ/jgZCMMZyklGKDMiu5WwETWUYRZPHoK//PIqaV3WfK/m319V6xdFHGU4gVM4Bx+uoQ530IAmMBjBM7zCm6OcF+fd+Vi0lpxi5hj+wPn8AfixjXM=</latexit>
(novel molecules for drugs j <latexit sha1_base64="jWfbUWAOv5Km2Y6nWW9FZ8Jfyxg=">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</latexit>

j and materials)
<latexit sha1_base64="IvN7ZIjsPg5Gkwiu2YyISSTnuhw=">AAAB6nicbVBNS8NAEJ34WetX1aOXxSJ4kJKIoMeCF48V7Qe0oWy2m3btZhN2J0IJ/QlePCji1V/kzX/jps1BWx8MPN6bYWZekEhh0HW/nZXVtfWNzdJWeXtnd2+/cnDYMnGqGW+yWMa6E1DDpVC8iQIl7ySa0yiQvB2Mb3K//cS1EbF6wEnC/YgOlQgFo2il+8dyuV+pujV3BrJMvIJUoUCjX/nqDWKWRlwhk9SYrucm6GdUo2CST8u91PCEsjEd8q6likbc+Nns1Ck5tcqAhLG2pZDM1N8TGY2MmUSB7Ywojsyil4v/ed0Uw2s/EypJkSs2XxSmkmBM8r/JQGjOUE4soUwLeythI6opQ5tOHoK3+PIyaV3UPLfm3V1W6+dFHCU4hhM4Aw+uoA630IAmMBjCM7zCmyOdF+fd+Zi3rjjFzBH8gfP5AzNqjQA=</latexit>

2 3
messages
Userj 4 images 5 2 Rd
videos
Social networks j
<latexit sha1_base64="JKJzsU+Q7PTS362hcrFlMMBdKDE=">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</latexit>

(Advertisement/ 2 3
a1
recommendation) 6 7
ROIi 4 ... 5 2 RT
aT i
<latexit sha1_base64="6N/ezo2rLj9wAqtHPmFaQH09OaE=">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</latexit>

i
<latexit sha1_base64="RlAIlE5MorzaIuoAsYZUvwE4eUI=">AAAB6nicbVBNS8NAEJ3Urxq/qh69LBbBg5REBD0WvHisaD+gDWWznbRLN5uwuxFK6U/w4kERr/4ib/4bN20O2vpg4PHeDDPzwlRwbTzv2ymtrW9sbpW33Z3dvf2DyuFRSyeZYthkiUhUJ6QaBZfYNNwI7KQKaRwKbIfj29xvP6HSPJGPZpJiENOh5BFn1Fjpgbtuv1L1at4cZJX4BalCgUa/8tUbJCyLURomqNZd30tNMKXKcCZw5vYyjSllYzrErqWSxqiD6fzUGTmzyoBEibIlDZmrvyemNNZ6Eoe2M6ZmpJe9XPzP62YmugmmXKaZQckWi6JMEJOQ/G8y4AqZERNLKFPc3krYiCrKjE0nD8FffnmVtC5rvlfz76+q9YsijjKcwCmcgw/XUIc7aEATGAzhGV7hzRHOi/PufCxaS04xcwx/4Hz+ADHkjP8=</latexit>
<latexit

Cerebral
ij A 2 R+
Brain
<latexit sha1_base64="f8VGxI+5fMbgb8XS6gM2gFgsuvQ=">AAAB63icbVBNS8NAEJ3Urxq/qh69LBbBg5REBD0WvHisYD+gDWWz3bRrdzdhdyKU0r/gxYMiXv1D3vw3Jm0O2vpg4PHeDDPzwkQKi5737ZTW1jc2t8rb7s7u3v5B5fCoZePUMN5ksYxNJ6SWS6F5EwVK3kkMpyqUvB2Ob3O//cSNFbF+wEnCA0WHWkSCUcwl8ei6/UrVq3lzkFXiF6QKBRr9yldvELNUcY1MUmu7vpdgMKUGBZN85vZSyxPKxnTIuxnVVHEbTOe3zshZpgxIFJusNJK5+ntiSpW1ExVmnYriyC57ufif100xugmmQicpcs0Wi6JUEoxJ/jgZCMMZyklGKDMiu5WwETWUYRZPHoK//PIqaV3WfK/m319V6xdFHGU4gVM4Bx+uoQ530IAmMBjBM7zCm6OcF+fd+Vi0lpxi5hj+wPn8AfixjXM=</latexit>

j connectionij ij <latexit sha1_base64="JZumyvT1BaYqivo266804eV0UKQ=">AAACAnicbVBNS8NAEJ3Urxq/op7ES7AIBaEkIuix4sVjFfsBbQib7aZdu9mE3Y1QQvHiX/HiQRGv/gpv/hs3bQ7a+mDg8d4MM/OChFGpHOfbKC0tr6yuldfNjc2t7R1rd68l41Rg0sQxi0UnQJIwyklTUcVIJxEERQEj7WB0lfvtByIkjfmdGifEi9CA05BipLTkWweXfkbvJ2aP8l6E1DAIstuJf2Kapm9VnJozhb1I3IJUoEDDt756/RinEeEKMyRl13US5WVIKIoZ0StSSRKER2hAuppyFBHpZdMXJvaxVvp2GAtdXNlT9fdEhiIpx1GgO/Mz5byXi/953VSFF15GeZIqwvFsUZgyW8V2nofdp4JgxcaaICyovtXGQyQQVjq1PAR3/uVF0jqtuU7NvTmr1KtFHGU4hCOoggvnUIdraEATMDzCM7zCm/FkvBjvxsestWQUM/vwB8bnD0E6le8=</latexit>

<latexit sha1_base64="IvN7ZIjsPg5Gkwiu2YyISSTnuhw=">AAAB6nicbVBNS8NAEJ34WetX1aOXxSJ4kJKIoMeCF48V7Qe0oWy2m3btZhN2J0IJ/QlePCji1V/kzX/jps1BWx8MPN6bYWZekEhh0HW/nZXVtfWNzdJWeXtnd2+/cnDYMnGqGW+yWMa6E1DDpVC8iQIl7ySa0yiQvB2Mb3K//cS1EbF6wEnC/YgOlQgFo2il+8dyuV+pujV3BrJMvIJUoUCjX/nqDWKWRlwhk9SYrucm6GdUo2CST8u91PCEsjEd8q6likbc+Nns1Ck5tcqAhLG2pZDM1N8TGY2MmUSB7Ywojsyil4v/ed0Uw2s/EypJkSs2XxSmkmBM8r/JQGjOUE4soUwLeythI6opQ5tOHoK3+PIyaV3UPLfm3V1W6+dFHCU4hhM4Aw+uoA630IAmMBjCM7zCmyOdF+fd+Zi3rjjFzBH8gfP5AzNqjQA=</latexit>

connectivity 2 3
a1
(sMRI) ROIj 6 . 7 T
4 .. 5 2 R
aT j
<latexit sha1_base64="60abWUf96kIh3P+ESEWT65J66ZM=">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</latexit>

Brain analysis Functional

(Neuroscience/neuro-diseases) activations (fMRI)
Xavier Bresson 15
 16

Graph-structured data

Social networks Scene understanding

(Advertisement) Word relationships
(NLP)

Brain
connectivity
Recommender
systems (Amazon, =
Drug/Material Netflix)
(Neuroscience)
molecules
(Chemistry)
Graph
Gene Regulatory
Network

Neutrino
3D Meshes Transportation detection (High-
(Computer networks energy Physics)
Graphics)
Xavier Bresson 16
 17

Graph-structured data 2 3
6 7
hj = 4 ... 5 2 Rdv
j <latexit sha1_base64="uzkOcVO/X53iwhYbLz3tS4wrJBE=">AAAB7HicbVBNS8NAEJ3Ur1q/qh69LBbBU0mKoMeCF48VTFtoQ9lsJ+3azSbsboRS+hu8eFDEqz/Im//GTZuDtj4YeLw3w8y8MBVcG9f9dkobm1vbO+Xdyt7+weFR9fikrZNMMfRZIhLVDalGwSX6hhuB3VQhjUOBnXBym/udJ1SaJ/LBTFMMYjqSPOKMGiv5jxWLQbXm1t0FyDrxClKDAq1B9as/TFgWozRMUK17npuaYEaV4UzgvNLPNKaUTegIe5ZKGqMOZotj5+TCKkMSJcqWNGSh/p6Y0VjraRzazpiasV71cvE/r5eZ6CaYcZlmBiVbLooyQUxC8s/JkCtkRkwtoUxxeythY6ooMzafPARv9eV10m7UPbfu3V/Vmo0ijjKcwTlcggfX0IQ7aIEPDDg8wyu8OdJ5cd6dj2VrySlmTuEPnM8fnmCNLg==</latexit>

<latexit sha1_base64="Ec0krwKP/HHMr51ajNWvoKg1wVM=">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</latexit>

ij 2 3
A graph G is defined by : <latexit sha1_base64="Tr3beZzVs88HzT/+beb54y3QhJs=">AAAB7XicbVBNSwMxEJ34WetX1aOXYBE8ld0i6LHgxWMF+wHtUrJptk2bTZYkK5Sl/8GLB0W8+n+8+W/MtnvQ1gcDj/dmmJkXJoIb63nfaGNza3tnt7RX3j84PDqunJy2jUo1ZS2qhNLdkBgmuGQty61g3UQzEoeCdcLpXe53npg2XMlHO0tYEJOR5BGnxDqpzSdlh0Gl6tW8BfA68QtShQLNQeWrP1Q0jZm0VBBjer6X2CAj2nIq2LzcTw1LCJ2SEes5KknMTJAtrp3jS6cMcaS0K2nxQv09kZHYmFkcus6Y2LFZ9XLxP6+X2ug2yLhMUsskXS6KUoGtwvnreMg1o1bMHCFUc3crpmOiCbUuoDwEf/XlddKu13yv5j9cVxv1Io4SnMMFXIEPN9CAe2hCCyhM4Ble4Q0p9ILe0ceydQMVM2fwB+jzB2SMjaE=</latexit>

=
Vertices V i 6 7
eij = 4 ... 5 2 Rde
<latexit sha1_base64="5YP62cqhTEZE4/44WrXhRx4PZp4=">AAAB7HicbVBNS8NAEJ34WetX1aOXxSJ4KkkR9Fjw4rGCaQttKJvtpF262YTdjVBCf4MXD4p49Qd589+4aXPQ1gcDj/dmmJkXpoJr47rfzsbm1vbObmWvun9weHRcOznt6CRTDH2WiET1QqpRcIm+4UZgL1VI41BgN5zeFX73CZXmiXw0sxSDmI4ljzijxko+r1oMa3W34S5A1olXkjqUaA9rX4NRwrIYpWGCat333NQEOVWGM4Hz6iDTmFI2pWPsWyppjDrIF8fOyaVVRiRKlC1pyEL9PZHTWOtZHNrOmJqJXvUK8T+vn5noNsi5TDODki0XRZkgJiHF52TEFTIjZpZQpri9lbAJVZQZm08Rgrf68jrpNBue2/AeruutZhlHBc7hAq7AgxtowT20wQcGHJ7hFd4c6bw4787HsnXDKWfO4A+czx+c2I0t</latexit>
<latexit

Edges E
Adjacency matrix A 2 3
<latexit sha1_base64="xxSF2mjlG89TsfNchohEXBU/le8=">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</latexit>

6 7
hi = 4 ... 5 2 Rdv

Graph features : G = (V, E, A)

<latexit sha1_base64="rsFij4wT42FKMyazBZXx3zauZAM=">AAACAHicbVDLSsNAFL3xWesr6sKFm8EiVCglKYJuhIqILivYB7ShTKaTdujkwcxEKCEbf8WNC0Xc+hnu/BsnbRbaeuDC4Zx7ufceN+JMKsv6NpaWV1bX1gsbxc2t7Z1dc2+/JcNYENokIQ9Fx8WSchbQpmKK004kKPZdTtvu+Drz249USBYGD2oSUcfHw4B5jGClpb552POxGhHMk9v0styq3FSuTosafbNkVa0p0CKxc1KCHI2++dUbhCT2aaAIx1J2bStSToKFYoTTtNiLJY0wGeMh7WoaYJ9KJ5k+kKITrQyQFwpdgUJT9fdEgn0pJ76rO7Nz5byXif953Vh5F07CgihWNCCzRV7MkQpRlgYaMEGJ4hNNMBFM34rICAtMlM4sC8Gef3mRtGpV26ra92elei2PowBHcAxlsOEc6nAHDWgCgRSe4RXejCfjxXg3PmatS0Y+cwB/YHz+ANOBk+A=</latexit>
<latexit sha1_base64="QWPKAQ5W09RWj9lneydNzjG7eKs=">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</latexit>

Node features : hi, hj (atom type) V = {1, .., n}

<latexit sha1_base64="uyVdbziOGHiCtZZ2/+CXtVUELUQ=">AAAB+XicbVBNS8NAEJ3Ur1q/oh69LBbBQwlJEfQiFLx4rGBroQlls920SzebsLsplNB/4sWDIl79J978N27aHLT1wcDjvRlm5oUpZ0q77rdV2djc2t6p7tb29g8Oj+zjk65KMklohyQ8kb0QK8qZoB3NNKe9VFIch5w+hZO7wn+aUqlYIh71LKVBjEeCRYxgbaSBbXdv/dxrOE5D+POawcCuu467AFonXknqUKI9sL/8YUKymApNOFaq77mpDnIsNSOczmt+pmiKyQSPaN9QgWOqgnxx+RxdGGWIokSaEhot1N8TOY6VmsWh6YyxHqtVrxD/8/qZjm6CnIk001SQ5aIo40gnqIgBDZmkRPOZIZhIZm5FZIwlJtqEVYTgrb68TrpNx3Md7+Gq3mqWcVThDM7hEjy4hhbcQxs6QGAKz/AKb1ZuvVjv1seytWKVM6fwB9bnD5twkPk=</latexit>

Edge features : eij (bond type)

Graph features : g (molecule energy)
2 3
6 7
g = 4 ... 5 2 Rdg
<latexit sha1_base64="bl0qxLKrClBChsCfAREYN8sxcnw=">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</latexit>

Xavier Bresson 17
 18

Outline

The Deep Learning (DL) revolution

Limitations of DL
Graph-Structured data
Graph Neural Networks (GNNs)
GNN case studies
GNN for industry
GNN libraries
Conclusion

Xavier Bresson 18
 19

Graph Neural Networks (GNNs)

Why generalizing Deep Learning to graphs is hard ?

Irregular grid/topology: the size of graphs varies and the number of neighbors changes
No node ordering: no coordinates/directions on graphs like up, down, right or left
Lots of symmetries:
Permutation equivariance/invariance
Isomorphic nodes, i.e. nodes with the same neighborhood structures
Dynamic topology: node/edge can appear and be discarded
Node-level prediction
Edge-level prediction

GNNs

+
+ f(G)

G
Graph-level prediction
Xavier Bresson 19
 20

GNNs in research
One of the hottest machine learning topics since 2021 !
Standard toolkit for analyzing and learning from data on graphs
Flexible to adapt to complex data structure and combine distinct modalities

ICLR

Xavier Bresson 20
 21

My research

Design minimalist and mathematically sound GNNs for a wide range of applications
Co-inventor of Graph Convolutional Networks in 2016 (10th most cited paper at NeurIPS)
Popularized GNNs w/ Michael Bronstein (Oxford), Yann LeCun (NYU/Meta/Turing)
with tutorials at NeurIPS’17, CVPR’17, SIAM AM’18
Developed with Yoshua Bengio’20’21 (MILA/Turing) and Yann LeCun’22’23 new classes of GNNs
Invited speaker at KDD’23, NeurIPS’22, AAAI’21, KDD’21’23, ICLR’20, ICML’20 (AI conferences)
Conference/workshop organizer at LoG’23’22, UCLA’23’21’19’18, ICLR’22, MLSys’21
Awarded the US$2M NRF Fellowship in 2017 (largest individual grant in Singapore)
16,000+ citations, 60+ articles, 70,000+ YouTube views, 8,000+ Twitter followers

NeurIPS’17 CVPR’17
1,000-2,000 500-1,000
participants participants

Xavier Bresson 21
 22

Outline

The Deep Learning (DL) revolution

Limitations of DL
Graph-Structured data
Graph Neural Networks (GNNs)
GNN case studies
GNN for industry
GNN libraries
Conclusion

Xavier Bresson 22
 23

GNN case studies

Chip design (Google)

Resource management
Scene reasoning (Meta)
Recommendation (UberEats/ Pinterest/ Alibaba)
Fake news detection
Finance
Natural Language Processing
Knowledge graphs (Amazon)
Transportation (Google Map ETA)
Autonomous driving (NVIDIA)
Protein & drug design (Google DeepMind/ Microsoft/ AstraZeneca)
Energy physics & simulations (Google DeepMind)
Code bug detection
Genomics

Xavier Bresson 23
 24

GNNs for chip design

https://ai.googleblog.com/2020/04/chip-design-with-deep-reinforcement.html

https://www.theverge.com/2021/6/10/22527476/google-machine-learning-chip-design-tpu-floorplanning

Xavier Bresson 24
 25

GNNs for resource allocation

Operations Research/Combinatorial
Optimization s.a. assignment, routing,
planning, supply chain, scheduling are used
every day in revenue management, Hardware pieces placement Uber taxis allocation
transportation, manufacturing, supply chain,
public policy, hardware design, etc.

Most OR problems are NP-hard.

Traditional OR solvers are hand-crafted

algorithms with years of research work and
https://towardsdatascience.com/can-transformers-solve-this-90-year-
significant specialized knowledge. old-classic-computer-science-problem-better-than-human-
8ef78bb2e9d8

DL can learn universal high-quality

algorithms to solve OR problems with GNNs
and RL.

Amazon warehouse
Xavier Bresson management 25
 26

GNNs for scene reasoning

xxx

https://deepmind.com/blo
g/article/neural-approach-
relational-reasoning

Xavier Bresson 26
 27

GNNs for autonomous driving

xxx

https://slideslive.com/38930570/graph-
neural-networks-for-selfdriving

Xavier Bresson 27
 28

GNNs for autonomous driving

Computer Vision with visual common sense

Reasoning on scene graph

A. Karpathy (Tesla AI
Autopilot Director)
https://youtu.be/2PpN
mSdFP7Q?t=3369

Scene graph understanding

Xavier Bresson 28
 29

GNNs for recommendation

xxx

https://eng.uber.com/uber-eats-graph-learning

https://medium.com/pinterest-engineering/pinsage-a-new-graph-convolutional-
neural-network-for-web-scale-recommender-systems-88795a107f48
Xavier Bresson 29
 30

GNNs for fake news detection

Social networks :
Fake News (2016 US Presidential Election)
/(human) adversarial attacks

Facebook network
Twitter network

Xavier Bresson 30
 31

GNNs for finance

Financial networks model financial entities and their relationships.
Applications :

client/bank/company
$100, date

Credit: You et-al, Arxiv 2021

Xavier Bresson 31
 32

GNNs for knowledge graphs

xxx
Knowledge graphs represent large-
scale data information in the form of
triples.
Triples are two entities and their
type of relationship.

https://www.amazon.science/blog/combining-knowledge-
graphs-quickly-and-accurately

Xavier Bresson 32
 33

GNNs for knowledge graphs

Reasoning on KGs can provide critical new information.
Predict drugs that target proteins associated to Covid-19.
Where did Canadian Turing Award winners graduate?
Limitations of non-ML reasoning on KG :
Credit: Amazon Neptune ML
KGs are incomplete, noisy.
Answer query on large graphs can become intractable with standard graph search
algorithms.
GNNs can represent complex queries in a continuous embedding space.
Answer is given by the closest entity (w.r.t. the Euclidean distance).

First-order logic Query

Formula

Xavier Bresson Credit: Hamilton et al, NeurIPS 2018 33

 34

GNNs for NLP

xxx

https://openai.com/blog/better-language-models

https://thegradient.pub/transformers-are-graph-neural-networks
Transformers are (fully-connected) Graph Neural Networks.
Xavier Bresson 34
 35

Limitation of Transformers

Impressive results at large-scale

2020 GPT-3, 175B parameters, 285,000 CPUs, 10,000 GPUs, 400Gb/sec network
connectivity, 500 billion tokens, US$12 Million to train
Transformers capture dynamic word representation depending on the context.
Example : The vase broke. The news broke. Sandy broke the world record. Sandy
broke the law. We broke even. The burglar broke into the house. Etc.
DL has not yet reached human performance (no common sense).

What is missing to get to human’s level?

More data? Yes but not sufficient.
Reasoning, but with what inductive bias?
Knowledge graphs

Credit: Antoine Bosselut, EPFL

Xavier Bresson 35
 36

GNNs for NLP

Transformers vs GNNs with knowledge graphs

Credit: Yasunaga et al, NAACL, 2021

Xavier Bresson 36
 37

GNNs for transportation

https://deepmind.com/blog/article/traffic-prediction-with-advanced-
graph-neural-networks
Xavier Bresson 37
 38

GNNs for protein folding

Graph Transformers

https://deepmind.com/blog/article/alphafold-a-solution-to-a-50-year-
old-grand-challenge-in-biology

Xavier Bresson 38
 39

GNNs for protein function & interaction

xxx

Xavier Bresson 39
 40

GNNs for drug design

https://www.sciencedirect.com/science/article/pii/S0092867420301021

Xavier Bresson 40
 41

GNNs for energy physics

xxx

https://news.fnal.gov/2020/09/the-next-big-thing-the-
use-of-graph-neural-networks-to-discover-particles

Xavier Bresson 41
 42

GNNs for energy physics

High-energy physics with neutrino detection (hard to detect
because they have a very small chance of interacting with
regular matter).

IceCube Neutrino Observatory

Xavier Bresson 5,160 sensors 42

 43

GNNs for physics simulation

xxx

https://deepmind.com/research/open-
source/Learning-to-Simulate-Complex-Physics-with-
Graph-Networks

Xavier Bresson 43
 44

GNNs for code bug detection

Experiments conducted on a large dataset
with 4.9M methods in 92 different project
versions show that GNNs have a relative
improvement up to 160% on F-score when
comparing with the state-of-the-art bug
detection approaches.

https://dl.acm.org/doi/pdf/10.1145/3360588

Xavier Bresson 44
 45

GNNs for genomics

Biological
sample

+
We proposed to use GNN to replace human
heuristics ⇒ AI-based genome assembler Genome

Advantage : Solve genome assembly

independently of any type of sequencing (Unordered) reads

machine and no hand-crafting of genome

assemblers.
Overlapping reads
Given a state-of-the-art genome assembler
(Raven), we demonstrated that learned + Graph construction
(Overlap phase)
heuristics with GNN outperforms human Assembly graph
engineered rules. (directed graph)

Longest path /
Human genome

2001 2022
https://arxiv.org/pdf/2206.00668.pdf
Xavier Bresson 45
 46

Outline

The Deep Learning (DL) revolution

Limitations of DL
Graph-Structured data
Graph Neural Networks (GNNs)
GNN case studies
GNN for industry
GNN libraries
Conclusion

Xavier Bresson 46
 47

GNNs for industry

Market potential has a large landscape :

GNN is a general technology that can be applied to several tasks
Finance: fraud/anomaly detection, forecast prediction for e.g. sales, investments, credit risk
Recommendation: better targeted ads, improved user experience/loyalty
Knowledge graphs: enhanced CRM, query relationships/interactions with company and
customers
Healthcare: drug design, new diagnostic tools for doctors s.a. brain analysis
Robotics: better 3D point representation, planning and reasoning
NLP: improved Q&A chatbot with contextual graphs
Resource management: supply chain and warehouse/inventory optimization
Transportation: more accurate and dynamic delivery time
Etc.

Xavier Bresson 47
 48

GNNs for industry

Recent technology

Kumo : https://kumo.ai
Start-up raised $18.5 Million in July 2022
A co-funder is Jure Leskovec (Stanford)

Amazon Neptune:
https://aws.amazon.com/neptune/machine-learning

Xavier Bresson 48
 49

GNN pipeline

Collect data (user/customer features, product features, etc.)

Generate graphs from data relationships and features

Train/validate GNN for given predictive tasks

Cloud storage, security, computing and deployment

Customers query an API to get predictions

Easy to use for non-expert DL/GNN customers

Xavier Bresson 49
 50

Outline

The Deep Learning (DL) revolution

Limitations of DL
Graph-Structured data
Graph Neural Networks (GNNs)
GNN case studies
GNN for industry
GNN books and libraries
Conclusion

Xavier Bresson 50
 51

Books
Graph Representation Learning Book, Springer, 2020
William Hamilton, McGill University
https://link.springer.com/book/10.1007/978-3-031-01588-5
https://github.com/RHxW/CV-DL-Docs/blob/master/GRL_Book.pdf

Geometric Deep Learning: Grids, Groups, Graphs, Geodesics, and Gauges,

2021
Michael M. Bronstein (Oxford), Joan Bruna (NYU), Taco Cohen
(Qualcomm), Petar Veličković (DeepMind)
https://arxiv.org/pdf/2104.13478.pdf

Graph Neural Networks: Foundations, Frontiers, Applications, Springer,

2022
Lingfei Wu (Pinterest), Peng Cui (Tsinghua), Jian Pei, (Duke), Liang
Zhao (Emory University)
https://graph-neural-networks.github.io (English and Chinese versions)
Xavier Bresson 51
 52

Libraries
Amazon DGL (Deep Graph Library)
First released in Dec 2018 (11k+ stars)
PyTorch / TensorFlow / MxNet
https://www.dgl.ai

Kumo PyG (PyTorch Geometric)

First released in Mar 2019 (16k+ stars)
PyTorch
https://www.pyg.org

TensorFlow GNNs
First released in Nov 2021 (1k+ stars)
https://github.com/tensorflow/gnn
Xavier Bresson 52
 53

Outline

The Deep Learning (DL) revolution

Limitations of DL
Graph-Structured data
Graph Neural Networks (GNNs)
GNN case studies
GNN for industry
GNN libraries
Conclusion

Xavier Bresson 53
 54

Conclusion

Graph Neural Networks are a breakthrough in Machine Learning.

GNNs have become the standard toolkit for analyzing graph-structured data.
They generalize CNNs/RNNs/Transformers from grids/sequences to complex relational
data structures.
Graphs are everywhere because everything is connected.
GNNs are highly flexible and have been/will be applied to a large variety of applications.
Large-scale training (linear complexity) with distributed computing.
Supervised, reinforcement and self-supervised training.
GNNs will boost business analytics.

Xavier Bresson 54
 55

Tentative Lectures
Introduction to Graph Machine Learning Part 3 : GML with deep feature learning,
Part 1: GML without feature learning a.k.a. GNNs (after 2016)
(before 2014) Graph Convolutional Networks
Introduction to Graph Science (spectral and spatial)
Graph Analysis Techniques without Weisfeiler-Lehman GNNs
Feature Learning Graph Transformer & Graph
Graph clustering ViT/MLP-Mixer
Classification Benchmarking GNNs
Recommendation Molecular science and generative GNNs
Dimensionality reduction GNNs for combinatorial optimization
Part 2 : GML with shallow feature learning GNNs for recommendation
(2014-2016) GNNs for knowledge graphs
Shallow graph feature learning Integrating GNNs and LLMs

Xavier Bresson 55
 56

Questions?

Xavier Bresson 56