Recurrent Graph Neural Networks

Giovanni Pellegrini1,2,3
SML1 Lab, University of Trento, Italy
TIM2
EIT DIGITAL3

1
01 GNNs so far

xA

xB

xE

Neural Networks
Permutation invariant
Aggregation

Sum
Average
Max

2
02 Recurrent VS Convolutional

3
02 Recurrent VS Convolutional

4
02 Recurrent VS Convolutional

5
02 Recurrent VS Convolutional

6
02 Recurrent VS Convolutional

7
02 Recurrent VS Convolutional

8
 GNNs so far 01

02
Recurrent VS
Convolutional

TABLE OF 04 Gated Graph NN

The Graph Neural
Network Model 03 CONTENTS
05
Gated Graph Sequence
NN

06 PyG Tutorial

9
 03 Graph Neural Network Model1

● Pioneer work on Graph Neural Network

● Diffusion mechanism (convolution)
● General framework for graph processing

1
Scarselli, et al., The graph neural network model, IEEE Transactions on Neural Networks, 2009
10
03 Graph Neural Network Model

11
03 Graph Neural Network Model

12
03 Graph Neural Network Model

13
03 Graph Neural Network Model

14
03 Graph Neural Network Model

15
03 Graph Neural Network Model
= edges connected to v

= neighbours of v

16
03 Graph Neural Network Model
= edges connected to v

= neighbours of v

17
03 Graph Neural Network Model
= edges connected to v

= neighbours of v

18
03 Graph Neural Network Model
= edges connected to v

= neighbours of v

learnable parameters

19
03 Graph Neural Network Model
= edges connected to v

= neighbours of v

learnable parameters

= transition function

= output function

20
 03 Graph Neural Network Model
Repeated forward of the transition function create an
encoding network

Image taken from the original publication

21
 03 Graph Neural Network Model
Goal: converge to a unique solution for xv and ov

22
 03 Graph Neural Network Model
Goal: converge to a unique solution for xv and ov

If the transition function is a contraction mapping,

there exists a fixed point solution

23
 03 Graph Neural Network Model
Goal: converge to a unique solution for xv and ov

If the transition function is a contraction mapping,

there exists a fixed point solution

If is a NN, to ensure the contraction mapping a

penalty based on the norm of the Jacobian is added to
the loss function

24
03 Graph Neural Network Model

25
03 Graph Neural Network Model

26
03 Graph Neural Network Model

27
03 Graph Neural Network Model

Almeida - Pineda
algorithm

28
 04 Gated Graph Neural Network2

● Adaptation of the GNNM

● Uses GRU (Gated Recurrent Units) as transition function
● Iterate over T timesteps (instead of until convergence)
● Uses BTT (BackProp Through Time) to compute gradient

1
Li et al, Gated graph sequence neural networks, ICLR, 2015.
29
 04 Gated Graph Neural Network
Node annotations: node embeddings with additional
information

Propagation Model:

30
 04 Gated Graph Neural Network
Node annotations: node embeddings with additional
information
Node Node
state annotation Propagation Model:

31
 04 Gated Graph Neural Network
Node annotations: node embeddings with additional
information
Node Node
state annotation Propagation Model:

32
 05 Gated Graph Sequence Neural Network
What if we want to produce sequences of output values?

33
 05 Gated Graph Sequence Neural Network
What if we want to produce sequences of output values?

GGSNN : multiple GGNNs operate in sequence to

produce

34
 05 Gated Graph Sequence Neural Network
What if we want to produce sequences of output values?

GGSNN : multiple GGNNs operate in sequence to

produce

Computes X(k+1) from X(k)

Computes o(k) from X(k)

35
 05 Gated Graph Sequence Neural Network
What if we want to produce sequences of output values?

GGSNN : multiple GGNNs operate in sequence to

produce

Computes X(k+1) from X(k)

Computes o(k) from X(k)

implements both the transition and output function!

36
 05 Gated Graph Sequence Neural Network

k = length of the sequence

t = timesteps

37
 06 PyG tutorial

Let’s go to the Jupyter notebook….

38