Recently, number of papers re-visited this problem of generalizing neural networks to work on arbitrarily structured graphs, some achieving promising results in domains that have previously been dominated by other shallower algorithms. While Graph convolutions are generalization of spacial convolutions, and easiest to define in spectral domain, General Fourier transform used to represent them scales poorly with size of data. Therefore, first order approximation in Fourier-domain used to obtain efficient linear-time graph-CNNs. Those scales poorly with size of data. due to that, the expressiveness power of the proposed graph convolutional networks is severely impoverished. Another approach for learning graph representations requires the repeated application of contraction maps as propagation functions until node representations reach a stable fixed point. We combine those approaches and propose a recurrent version of Relational Graph Convolution networks, we then proceed to construct two models, Recurrent Variational Graph AutoEncoder and Recurrent Graph Convolution Regressor and show that for Ethereum Blockchain transaction graph we outperform the traditional Graph Convolution Network at predicting future movments of the corresponding tradable asset: Ether.

1. Recurrent Graph Convolution Networks for
Forecasting Ethereum prices
ICCS 2018
In collaboration with

2. tl;dr: We extended Graph Convolution
Networks to be Recurrent over time.

3. What is Ethereum
- A 100% open source platform to build
and distribute decentralized
applications
- No middle men
- Social sites, Financial systems,
Voting mechanisms, Games,
Reputation Systems
- 100% peer to peer, censorship proof
- Also a Tradable Asset.
RECURRENT GRAPH NEURAL NETWORKS

4. Experiment Setting

5. EXPERIMENT SETTING
Time 0 900 1800 2700 3600 4500 5400 6300 7200 8100 9000 9900 10800 11700 12600 13500 14400 15300 16200 17100 18000 18900 19800 … … … … … … … …
Batch - 1
Batch - 2
Batch - 3
Batch - 4
Batch - 5
Batch - 6
Batch - 7
Batch - 8
Batch - 9
…
…
…
Optimization Window Unseen
Vector – 60 Min lagged prices
Ground Truth – ETH Future 5 Min Prices
Batch – Training: 240 Vectors Test: 90 Vectors
In Sample
Training set: 28.02.18 - 13.05.18
Test set: 13.05.18 – 29.05.18
Out of sample
5 (min)60 (min)
Vector Structure
Optimization Window Unseen
Optimization Window Unseen
Optimization Window Unseen
Optimization Window Unseen
Optimization Window Unseen
Optimization Window Unseen
Optimization Window Unseen
Optimization Window Unseen
RECURRENT GRAPH NEURAL NETWORKS

6. DEEP LEARNING SUPERIORITY
96.92%
Deep Learning
94.9%
Human
ref: http://www.image-net.org/challenges/LSVRC/
RECURRENT GRAPH NEURAL NETWORKS

7. GRADIENT DESCENT
𝐸 = Error of the network
𝑤𝑡 = 𝑤𝑡−1 − 𝛾
𝜕𝐸
𝜕𝑤
𝑊 = Weight matrix representing the filters
RECURRENT GRAPH NEURAL NETWORKS

8. BackPropagation
Legend
𝑥0
𝑓0(𝑥0, 𝑤0)
𝑓1(𝑥1, 𝑤1)
𝑓2(𝑥2, 𝑤2)
𝑓𝑛 𝑥 𝑛, 𝑤 𝑛 = ො𝑦
𝑓𝑛−1(𝑥 𝑛−1, 𝑤 𝑛−1)
𝑓𝑛−2(𝑥 𝑛−2, 𝑤 𝑛−2)
𝑤0
𝑤1
𝑤 𝑛
𝑤 𝑛−1
𝐸 = 𝑙 ො𝑦, 𝑦𝑦
𝑙 ො𝑦, 𝑦 - Loss Function
𝑥0 - Features Vector
𝑥𝑖 - Output of 𝑖 layer
𝑤𝑖 - Weights of 𝑖 layer
𝑦 – Ground Truth
ො𝑦 – Model Output
𝐸 – Loss Surface
𝜕𝐸
𝜕𝑥 𝑛
=
𝜕𝑙 ො𝑦, 𝑦
𝜕𝑥 𝑛
𝜕𝐸
𝜕𝑤 𝑛
=
𝜕𝐸
𝜕𝑥 𝑛
𝜕𝑓𝑛 𝑥 𝑛−1, 𝑤 𝑛
𝜕𝑤 𝑛
𝜕𝐸
𝜕𝑥 𝑛−1
=
𝜕𝐸
𝜕𝑥 𝑛
𝜕𝑓𝑛 𝑥 𝑛−1, 𝑤 𝑛
𝑥 𝑛−1
𝑓– Activation Function
𝜕𝐸
𝜕𝑥 𝑛−2
=
𝜕𝐸
𝜕𝑥 𝑛−1
𝜕𝑓𝑛−1 𝑥 𝑛−2, 𝑤 𝑛−1
𝑥 𝑛−2
𝜕𝐸
𝜕𝑤 𝑛−1
=
𝜕𝐸
𝜕𝑥 𝑛−1
𝜕𝑓𝑛 𝑥 𝑛−2, 𝑤 𝑛−1
𝜕𝑤 𝑛−1
…
…
𝐹𝑜𝑟𝑤𝑎𝑟𝑑𝑃𝑟𝑜𝑝𝑎𝑔𝑎𝑡𝑖𝑜𝑛
𝐵𝑎𝑐𝑘𝑃𝑟𝑜𝑝𝑎𝑔𝑎𝑡𝑖𝑜𝑛
1: Forward Propagation 2: Loss Calculation 3: Optimization
RECURRENT GRAPH NEURAL NETWORKS

9. CONVOLUTION
ඵ
−∞−∞
∞∞
𝑓 𝜏1, 𝜏2 ∙ 𝑔 𝑥 − 𝜏1, 𝑦 − 𝜏2 𝑑𝜏1 𝑑𝜏2
𝑓 𝑥, 𝑦 𝑔 𝑥, 𝑦 𝑓 ∗ 𝑔
RECURRENT GRAPH NEURAL NETWORKS

10. ConvNet
𝑥Input
ො𝑦
Class
Convolution
&
Maxpooling
Convolution
&
Maxpooling
Convolution
&
Maxpooling
Fully Connected
RECURRENT GRAPH NEURAL NETWORKS

11. F1RMSE PnL(%)
Results: 1D-ConvNet
0.9 0.58 -17.317
Results for out of sample simulated trading
Simple root mean square error F1-beta score (Harmonic mean of
precision and recall) taken as
classification decision where the
predicted price is greater then the
current price +15% transaction fee.
Profits and losses (percentage) for
out of sample trading.
Assuming 15% transaction fee.
RECURRENT GRAPH NEURAL NETWORKS

12. Recurrent Neural Network
-Memory Achieved through feedback
-Due to self multiplications, Feedback Weight matrix tend to explode or vanish.
-Solution: logistic gating mechanism
Keep
Gate
1.73
Write
Gate
Read
Gate
Input from
rest of RNN
Output to
rest of RNN
Input
Command Output
Cell Gate
RECURRENT GRAPH NEURAL NETWORKS

13. Recurrent Neural Network
𝒔 𝒕+𝟏𝒔 𝒕 𝒔 𝒕+𝟐…….. ……..Backpropagation
Through Time
Long Short Term
Memory
ቁ𝑓𝑖
𝑙+1
𝑡
= 𝜎𝑔( 𝜔 𝑓 𝑦𝑗
𝑙
𝜍 𝑡−1 + 𝜓 𝑓ℎ 𝑡−1 + 𝑏𝑓
ቁ𝜄𝑖
𝑙+1
𝑡
= 𝜎𝑔( 𝜔𝜄 𝑦𝑗
𝑙
𝜍 𝑡−1 + 𝜓𝜄ℎ 𝑡−1 + 𝑏𝜄
ቁ𝑜𝑖
𝑙+1
𝑡
= 𝜎𝑔( 𝜔 𝑜 𝑦𝑗
𝑙
𝜍 𝑡−1 + 𝜓 𝑜ℎ 𝑡−1 + 𝑏 𝑜
ቁ𝜍𝑖
𝑙+1
𝑡
= 𝑡𝑎𝑛ℎ( 𝜔𝜍 𝑦𝑗
𝑙
+ 𝜓𝜍ℎ 𝑡−1 + 𝑏𝜍
Forget Gate
Output Gate
Input Gate
Cell
RECURRENT GRAPH NEURAL NETWORKS

14. F1RMSE PnL(%)
Results: LSTM
0.1 0.42 -7.115
Results for out of sample simulated trading
Simple root mean square error F1-beta score (Harmonic mean of
precision and recall) taken as
classification decision where the
predicted price is greater then the
current price +15% transaction fee.
Profits and losses (percentage) for
out of sample trading.
Assuming 15% transaction fee.
RECURRENT GRAPH NEURAL NETWORKS

15. INPUT
BIDIRECTIONAL GRU
RESIDUAL
DIALATED
CONV1D
𝒈 𝒕+𝟏𝒈 𝒕 𝒈 𝒕+𝟐
𝒈 𝒕+𝟏𝒈 𝒕+𝟐 𝒈 𝒕
TRANSPOSE
AXIS=1
tanh
softmax
tanh
softmax
tanh
softmax
HARD ATTENTION

16. F1RMSE PnL(%)
Results: CNN-LSTM
0.05 0.53 -7.461
Results for out of sample simulated trading
Simple root mean square error F1-beta score (Harmonic mean of
precision and recall) taken as
classification decision where the
predicted price is greater then the
current price +15% transaction fee.
Profits and losses (percentage) for
out of sample trading.
Assuming 15% transaction fee.
RECURRENT GRAPH NEURAL NETWORKS

17. CNN?
RNN?
CNN-RNN?
…

18. DEEP LEARNING COMMON STRUCTURES
SUPERVISED UNSUPERVISED
Perceptron It is a type of linear classifier, a classification algorithm that makes its predictions based on a linear predictor function
combining a set of weights with the feature vector. The algorithm allows for online learning, in that it processes elements in the
training set one at a time.
RECURRENTFEED FORWARD
Feed Forward Network sometimes
Referred to as MLP, is a fully connected
dense model used as a simple
classifier.
Convolutional Network assume that
highly correlated features located close
to each other in the input matrix and
can be pooled and treated as one in the
next layer.
Known for superior Image classification
capabilities.
Simple Recurrent Neural Network is a
class of artificial neural network where
connections between units form a
directed cycle.
Hopfield Recurrent Neural Network It is
a RNN in which all connections are
symmetric. it requires stationary
inputs.
Long Short Term Memory Network
contains gates that determine if the
input is significant enough to
remember, when it should continue to
remember or forget the value, and when
it should output
Auto Encoder aims to learn a
representation (encoding) for a set of
data, typically for the purpose of
dimensionality reduction.
Restricted Boltzmann Machine can
learn a probability distribution over its
set of inputs..
Deep Belief Net is a composition of
simple, unsupervised networks such as
restricted Boltzmann machines ,where
each sub-network's hidden layer serves
as the visible layer for the next.
RECURRENT GRAPH NEURAL NETWORKS

19. Search Problems
=
=
=
RECURRENT GRAPH NEURAL NETWORKS

20. Markov Decision Process
Action: 𝒂 𝒕 Action: 𝒂 𝒕+𝟏
Reward : 𝒓 𝒕 Reward : 𝒓 𝒕+𝟏 Reward : 𝒓 𝒕+𝟐
𝒔 𝒕+𝟏𝒔 𝒕 𝒔 𝒕+𝟐…….. ……..
𝑆 ≔ {𝑠1, 𝑠2, 𝑠3, … 𝑠 𝑛}
𝐴 ≔ {𝑎1, 𝑎2, 𝑎3, … 𝑎 𝑛}
𝑇(𝑠, 𝑎, 𝑠𝑡+1)
𝑅(𝑠, 𝑎)
Set of states
Set of Actions
Reward Function
Transition Function
RECURRENT GRAPH NEURAL NETWORKS

21. Policy Search
𝒔 𝒕 𝝅
𝑸(𝒔, 𝒂)
𝑸(𝒔, 𝒂)
𝑸(𝒔, 𝒂)
𝑸(𝒔, 𝒂)
Policy Expected Reward
𝝅: 𝒔 → 𝒂
The goal will be to
Maximize the reward
RECURRENT GRAPH NEURAL NETWORKS

22. Reinforcement Learning
Observation
Action
Value – Maps state, action pair to expected future reward
𝑸 𝒔, 𝒂 ≈ 𝔼 𝑹 𝒕+𝟏 + 𝑹 𝒕+𝟐 + 𝑹 𝒕+𝟑 + … 𝑺𝒕 = 𝒔, 𝑨 𝒕 = 𝒂]
Optimal Value – Bellman Equation 1957
𝑸∗
𝒔, 𝒂 ≈ 𝔼 𝑹 𝒕+𝟏 𝑸∗
(𝑺𝒕+𝟏, 𝒃) 𝑺 𝒕 = 𝒔, 𝑨 𝒕 = 𝒂]
TD Algorithm – Watkins 1989
𝑸 𝒕+𝟏 𝑺𝒕, 𝑨 𝒕 = 𝑸 𝒕 𝑺 𝒕, 𝑨 𝒕 + 𝜶(𝑹 𝒕+𝟏 + γmax
𝑎
𝑸 𝒕 𝑺𝒕+𝟏, 𝑨 𝒕 − 𝑸 𝒕 𝑺𝒕, 𝑨 𝒕 ]
RECURRENT GRAPH NEURAL NETWORKS

23. Gets “Rewards” and Penalties based on
it’s success of producing a better
generation of models.
Father Model
Being built, compiled, evaluated and
stored for future reconstruction and
retraining by a Human.
Child Model
Deep
Reinforcement
Learning
Deep Meta Learning
RECURRENT GRAPH NEURAL NETWORKS

25. F1RMSE PnL(%)
Results: Deep Meta Learning
0.027 0.68 -3.2
Results for out of sample simulated trading
Simple root mean square error F1-beta score (Harmonic mean of
precision and recall) taken as
classification decision where the
predicted price is greater then the
current price +15% transaction fee.
Profits and losses (percentage) for
out of sample trading.
Assuming 15% transaction fee.
RECURRENT GRAPH NEURAL NETWORKS

26. Reward Shaping
Random Walk LSTM
RECURRENT GRAPH NEURAL NETWORKS

28. <hash> <hash>
<hash> <hash>
<hash> <hash>
<hash> <hash>
<hash> <hash>
<hash> <hash>
<Amount>
<Amount>
<Amount>
<Amount>
<Amount>
<Amount>
…
…
Blockchain Representation

29. Learning Graph Representations
Random Walks On Graphs
Perozzi et al., 2014
Spectral networks
Bruna et al., 2013
Marginalized kernels between labeled graphs
Kashima, 2013
Graph Neural Networks
Gori 2015
Convolutional Networks on Graphs for
Learning Molecular Fingerprints
Duvenaud, 2015
RECURRENT GRAPH NEURAL NETWORKS

30. Spectral Networks
Convolution are diagonalized in Fourier Domain:
𝒙 ∗ 𝒉 = 𝓕−𝟏 𝒅𝒊𝒂𝒈 𝓕𝒉 𝓕𝒙
Where
𝓕 𝒌,𝒍 = 𝒆
(
−𝟐𝝅𝒊(𝒌∙𝒍)
𝑵 𝒅 )
Fourier basis can be defined as the eigenbasis of
Laplacian operator:
∆𝒙 𝒖 = ෍
𝒋≤𝒅
𝝏 𝟐 𝒙
𝝏𝒖𝒋
𝟐
(𝒖)
RECURRENT GRAPH NEURAL NETWORKS

31. Laplacian
𝑓
𝑓′
𝑓′′
𝛻𝑓
𝛻 ∙
𝐺𝑟𝑎𝑑𝑖𝑒𝑛𝑡
𝐷𝑖𝑣𝑒𝑟𝑔𝑒𝑛𝑐𝑒
RECURRENT GRAPH NEURAL NETWORKS

32. Graph Laplacian
RECURRENT GRAPH NEURAL NETWORKS

33. Graph Convolution
Spectral graph convolution
multiplication of a signal with a filter in the Fourier space of a graph.
Graph Fourier transform
multiplication of a graph signal 𝑋(i.e. feature vectors for every node)
with the eigenvector matrix 𝑈of the graph Laplacian 𝐿.
Graph Laplacian
can be easily computed from the symmetrically normalized graph adjacency
matrix ҧ𝐴: 𝐿 = 𝐼 − ҧ𝐴
Fourier basis of 𝑿 are Eigenvectors 𝑽 of 𝑳
RECURRENT GRAPH NEURAL NETWORKS

34. Spectral Networks
Convolution of Graph:
𝒙 ∗ 𝒉 𝒌 = 𝑽𝒅𝒊𝒂𝒈(𝒉)𝑽 𝑻
𝒙
RECURRENT GRAPH NEURAL NETWORKS

35. Translation Invariance?

36. Graphs Isomorphism
RECURRENT GRAPH NEURAL NETWORKS

37. ConvNet (LeNet5)
𝑥Input
ො𝑦
Class
Convolution
&
Maxpooling
Convolution
&
Maxpooling
Convolution
&
Maxpooling
Fully Connected
RECURRENT GRAPH NEURAL NETWORKS

38. ConvNet (LeNet)
𝑥Input
ො𝑦
Class
Convolution
&
Maxpooling
Convolution
&
Maxpooling
Convolution
&
Maxpooling
Fully Connected
Classifier
Representation
Learning
RECURRENT GRAPH NEURAL NETWORKS

39. Representation Bank
Give me the best
representation
for “cat”
the best
representation
for “cat”
Cat
RECURRENT GRAPH NEURAL NETWORKS

40. Single CNN layer with 3X3 filter
Convolutional Neural Network
2D
1D
RECURRENT GRAPH NEURAL NETWORKS

41. Single CNN layer with 3X3 filter
Euclidian Space Convolution
Update for a single pixel
-Transform neighbors individually 𝑤𝑖
(𝑙)
ℎ𝑖
(𝑙)
-Add everything up σ𝑖 𝑤𝑖
(𝑙)
ℎ𝑖
(𝑙)
-Add everything up ℎ0
(𝑙+1)
= 𝜎(σ𝑖 𝑤𝑖
(𝑙)
ℎ𝑖
(𝑙)
)
𝒉 𝟎
𝒉 𝟐𝒉 𝟏 𝒉 𝟑
𝒉 𝟒
𝒉 𝟓𝒉 𝟔𝒉 𝟕
𝒉 𝟖
𝑤7
𝑤8
𝑤1 𝑤2
𝑤3
𝑤4
𝑤5𝑤6
RECURRENT GRAPH NEURAL NETWORKS

42. Euclidian Space Convolution
𝒉 𝟎
𝒉 𝟐𝒉 𝟏 𝒉 𝟑
𝒉 𝟒
𝒉 𝟓𝒉 𝟔𝒉 𝟕
𝒉 𝟖
𝑤7
𝑤8
𝑤1 𝑤2
𝑤3
𝑤4
𝑤5𝑤6
ℎ0
(𝑙+1)
= 𝜎(෍
𝑖
𝑤𝑖
(𝑙)
ℎ𝑖
(𝑙)
)
RECURRENT GRAPH NEURAL NETWORKS

43. Graph Convolution as Message Passing
𝒉 𝟎
(𝒍+𝟏)
= 𝝈(𝒉 𝟎
(𝒍)
𝒘 𝟎
(𝒍)
෍
𝒊∈𝝒
𝟏
𝒄𝒊,𝒍
𝒘𝒋
(𝒍)
𝒉𝒋
(𝒍)
)
Propagation rule
𝒘 𝟎
𝒘 𝟏
𝒉𝒊
RECURRENT GRAPH NEURAL NETWORKS

44. def relational_graph_convolution(self, inputs):
features = inputs[0]
A = inputs[1:] # list of basis functions
# convolve
supports = list()
for i in range(support):
supports.append(K.dot(A[i], features))
supports = K.concatenate(supports, axis=1)
output = K.dot(supports, self.W)
𝒉 𝟎
(𝒍+𝟏)
= 𝝈(𝒉 𝟎
(𝒍)
𝒘 𝟎
(𝒍)
෍
𝒊∈𝝒
𝟏
𝒄𝒊,𝒍
𝒘𝒋
(𝒍)
𝒉𝒋
(𝒍)
)
RECURRENT GRAPH NEURAL NETWORKS

45. GRAPH CONVOLUTIONAL NETWORKS
ReLUReLU
Input
Features for nodes
𝑋 ∈ ℝ 𝑁∗𝐸
Adjacency matrix
containing all links መ𝐴
Embeddings
Representations that combine features of
neighborhood
Neighborhood size depends on number of
layers
RECURRENT GRAPH NEURAL NETWORKS

46. Problem
Embeddings are not optimized
For classification task!

47. GRAPH CONVOLUTIONAL NETWORKS
ReLUReLU
Input
Features for nodes
𝑋 ∈ ℝ 𝑁∗𝐸
Adjacency matrix
containing all links መ𝐴
Evaluate loss on labeled nodes only
ℒ = − ෍
𝐼∈𝑦 𝑖
෍
𝑓=𝐼
𝐹
𝑌𝐼𝑓ln(෍
𝑖
𝑒 𝑥 𝑖)
RECURRENT GRAPH NEURAL NETWORKS

48. EXAMPLE OF FORWARD PASS
𝑓( ) =
RECURRENT GRAPH NEURAL NETWORKS

49. Inits
SEMI-SUPERVISED CLASSIFICATION WITH
GRAPH CONVOLUTIONAL NETWORKS
Move
Nodes
https://github.com/tkipf/gcn
RECURRENT GRAPH NEURAL NETWORKS

50. Inits
SEMI-SUPERVISED CLASSIFICATION WITH
GRAPH CONVOLUTIONAL NETWORKS
Move
Nodes
https://github.com/tkipf/gcn
𝒉 𝟎
(𝒍+𝟏)
= 𝝈(෍
𝒊∈𝝒
𝟏
𝒄𝒊,𝒍
𝒘 𝟏
(𝒍)
𝒉𝒋
(𝒍)
)
RECURRENT GRAPH NEURAL NETWORKS

51. F1RMSE PnL(%)
Results: Graph Convolution
0.037 0.71 0.3
Results for out of sample simulated trading
Simple root mean square error F1-beta score (Harmonic mean of
precision and recall) taken as
classification decision where the
predicted price is greater then the
current price +15% transaction fee.
Profits and losses (percentage) for
out of sample trading.
Assuming 15% transaction fee.
RECURRENT GRAPH NEURAL NETWORKS

52. Temporal?

53. Recurrent Neural Networks Graph Convolution Networks
𝒘 𝟎
𝒘 𝟏
𝒉𝒊
𝒉 𝟎
(𝒍+𝟏)
= 𝝈(𝒉 𝟎
(𝒍)
𝒘 𝟎
(𝒍)
෍
𝒊∈𝝒
𝟏
𝒄𝒊,𝒍
𝒘𝒋
(𝒍)
𝒉𝒋
(𝒍)
)
ቁ𝑓𝑖
𝑙+1
𝑡
= 𝜎𝑔( 𝜔 𝑓 𝑦𝑗
𝑙
𝜍 𝑡−1 + 𝜓 𝑓ℎ 𝑡−1 + 𝑏𝑓
ቁ𝜄𝑖
𝑙+1
𝑡
= 𝜎𝑔( 𝜔𝜄 𝑦𝑗
𝑙
𝜍 𝑡−1 + 𝜓𝜄ℎ 𝑡−1 + 𝑏𝜄
ቁ𝑜𝑖
𝑙+1
𝑡
= 𝜎𝑔( 𝜔 𝑜 𝑦𝑗
𝑙
𝜍 𝑡−1 + 𝜓 𝑜ℎ 𝑡−1 + 𝑏 𝑜
ቁ𝜍𝑖
𝑙+1
𝑡
= 𝑡𝑎𝑛ℎ( 𝜔𝜍 𝑦𝑗
𝑙
+ 𝜓𝜍ℎ 𝑡−1 + 𝑏𝜍

54. Recurrent Neural Networks Graph Convolution Networks
𝒉 𝟎
(𝒍+𝟏)
= 𝝈(𝒉 𝟎
(𝒍)
𝒘 𝟎
(𝒍)
෍
𝒊∈𝝒
𝟏
𝒄𝒊,𝒍
𝒘𝒋
(𝒍)
𝒉𝒋
(𝒍)
)
ቁ𝑓𝑖
𝑙+1
𝑡
= 𝜎𝑔( 𝜔 𝑓 𝑦𝑗
𝑙
𝜍 𝑡−1 + 𝜓 𝑓ℎ 𝑡−1 + 𝑏𝑓
ቁ𝜄𝑖
𝑙+1
𝑡
= 𝜎𝑔( 𝜔𝜄 𝑦𝑗
𝑙
𝜍 𝑡−1 + 𝜓𝜄ℎ 𝑡−1 + 𝑏𝜄
ቁ𝑜𝑖
𝑙+1
𝑡
= 𝜎𝑔( 𝜔 𝑜 𝑦𝑗
𝑙
𝜍 𝑡−1 + 𝜓 𝑜ℎ 𝑡−1 + 𝑏 𝑜
ቁ𝜍𝑖
𝑙+1
𝑡
= 𝑡𝑎𝑛ℎ( 𝜔𝜍 𝑦𝑗
𝑙
+ 𝜓𝜍ℎ 𝑡−1 + 𝑏𝜍
൱𝑓𝑖
𝑙+1
𝑡
= 𝜎𝑔( 𝜔 𝑓 ෍
𝑟∈ℛ
൱෍
𝑗∈𝒩𝑖
𝑟
1
𝑐𝑖,𝑟
𝜃𝑟
𝑙
𝑦𝑗
𝑙
+ 𝜃0
𝑙
𝑦𝑖
𝑙
𝜍 𝑡−1 + 𝜓 𝑓ℎ 𝑡−1 + 𝑏𝑓
൱𝜄𝑖
𝑙+1
𝑡
= 𝜎𝑔( 𝜔𝜄 ෍
𝑟∈ℛ
൱෍
𝑗∈𝒩𝑖
𝑟
1
𝑐𝑖,𝑟
𝜃𝑟
𝑙
𝑦𝑗
𝑙
+ 𝜃0
𝑙
𝑦𝑖
𝑙
𝜍 𝑡−1 + 𝜓𝜄ℎ 𝑡−1 + 𝑏𝜄
൱𝑜𝑖
𝑙+1
𝑡
= 𝜎𝑔( 𝜔 𝑜 ෍
𝑟∈ℛ
൱෍
𝑗∈𝒩𝑖
𝑟
1
𝑐𝑖,𝑟
𝜃𝑟
𝑙
𝑦𝑗
𝑙
+ 𝜃0
𝑙
𝑦𝑖
𝑙
𝜍 𝑡−1 + 𝜓 𝑜ℎ 𝑡−1 + 𝑏 𝑜
൱𝜍𝑖
𝑙+1
𝑡
= 𝑡𝑎𝑛ℎ( 𝜔𝜍 ෍
𝑟∈ℛ
൱෍
𝑗∈𝒩𝑖
𝑟
1
𝑐𝑖,𝑟
𝜃𝑟
𝑙
𝑦𝑗
𝑙
+ 𝜃0
𝑙
𝑦𝑖
𝑙
+ 𝜓𝜍ℎ 𝑡−1 + 𝑏𝜍
Forget Gate
Output Gate
Input Gate
Cell

55. RECURRENT GRAPH CONVOLUTIONAL NETWORKS
RECURRENT GRAPH NEURAL NETWORKS
൱𝑓𝑖
𝑙+1
𝑡
= 𝜎𝑔( 𝜔 𝑓 ෍
𝑟∈ℛ
൱෍
𝑗∈𝒩𝑖
𝑟
1
𝑐𝑖,𝑟
𝜃𝑟
𝑙
𝑦𝑗
𝑙
+ 𝜃0
𝑙
𝑦𝑖
𝑙
𝜍 𝑡−1 + 𝜓 𝑓ℎ 𝑡−1 + 𝑏𝑓
൱𝜄𝑖
𝑙+1
𝑡
= 𝜎𝑔( 𝜔𝜄 ෍
𝑟∈ℛ
൱෍
𝑗∈𝒩𝑖
𝑟
1
𝑐𝑖,𝑟
𝜃𝑟
𝑙
𝑦𝑗
𝑙
+ 𝜃0
𝑙
𝑦𝑖
𝑙
𝜍 𝑡−1 + 𝜓𝜄ℎ 𝑡−1 + 𝑏𝜄
൱𝑜𝑖
𝑙+1
𝑡
= 𝜎𝑔( 𝜔 𝑜 ෍
𝑟∈ℛ
൱෍
𝑗∈𝒩𝑖
𝑟
1
𝑐𝑖,𝑟
𝜃𝑟
𝑙
𝑦𝑗
𝑙
+ 𝜃0
𝑙
𝑦𝑖
𝑙
𝜍 𝑡−1 + 𝜓 𝑜ℎ 𝑡−1 + 𝑏 𝑜
൱𝜍𝑖
𝑙+1
𝑡
= 𝑡𝑎𝑛ℎ( 𝜔𝜍 ෍
𝑟∈ℛ
൱෍
𝑗∈𝒩𝑖
𝑟
1
𝑐𝑖,𝑟
𝜃𝑟
𝑙
𝑦𝑗
𝑙
+ 𝜃0
𝑙
𝑦𝑖
𝑙
+ 𝜓𝜍ℎ 𝑡−1 + 𝑏𝜍
Forget Gate
Output Gate
Input Gate
Cell
ℎ 𝑡 ℎ 𝑡+1 ℎ 𝑡+2 ℎ 𝑡+3

56. F1RMSE PnL(%)
Results: Recurrent Graph Convolution
0.028 0.77 2.4
Results for out of sample simulated trading
Simple root mean square error F1-beta score (Harmonic mean of
precision and recall) taken as
classification decision where the
predicted price is greater then the
current price +15% transaction fee.
Profits and losses (percentage) for
out of sample trading.
Assuming 15% transaction fee.
RECURRENT GRAPH NEURAL NETWORKS

57. STRATEGY GRADIENT?
𝜕( 𝜃)
𝜕𝜃
−
𝜕( 𝜑)
𝜕𝜑
Returns Risk
RECURRENT GRAPH NEURAL NETWORKS

58. Graph Auto Encoders
𝑨 – input graph
𝒙 – input node
෡𝑨 – output graph
Useful for predicting connectivity links
RECURRENT GRAPH NEURAL NETWORKS

59. Recommender Systems
Users
Items
Graph
Representation
Graph
Prediction
Graph
AutoEncoder
RECURRENT GRAPH NEURAL NETWORKS

60. 𝜎
μ
Simulation
TRADING STRATEGY GRADIENTS
෍
𝑖=1
𝑛
𝜎𝑖
2
+ 𝜇𝑖
2
− log 𝜎𝑖 − 1
| ො𝑦 − 𝑦| 2
2
Action:
𝒂 𝒕+𝟏
RECURRENT GRAPH NEURAL NETWORKS

61. F1RMSE PnL(%)
Results: Recurrent Graph Auto Encoder
0.024 0.86 5.6
Results for out of sample simulated trading
Simple root mean square error F1-beta score (Harmonic mean of
precision and recall) taken as
classification decision where the
predicted price is greater then the
current price +15% transaction fee.
Profits and losses (percentage) for
out of sample trading.
Assuming 15% transaction fee.
RECURRENT GRAPH NEURAL NETWORKS

62. Conclusions
-Deep Learning works well on Euclidean data.
-Attempts to utilize DL for Non-Euclidean are starting to become viable.
-Reward shaping and drifted metrics are extremely misleading.
-After trying heavily we conclude that Aggregated data (prices) of Ethereum is
insufficient when trying to forecast behavior.
-We introduce a novel layer: Recurrent Graph Convolution and demonstrate
How this approach yield “tradable” results.
RECURRENT GRAPH NEURAL NETWORKS

63. FIN