Federated learning based_trafiic_flow_prediction.ppt

1
NIT Delhi
Federated Learning Based Traffic Flow
Prediction Using GCNN
Presented By :
MOH KHALID
Roll No. 192211009
M.Tech-CSE(Analytics)-NIT Delhi
May - 2020

Content
1. Introduction
2. Motivation
3. Objective
4. Problem Statement
5. Methodology
6. Result
7. Analysis
8. Scope of work
9. Conclusion Future Scope
10. References
2
NIT Delhi

Introduction
1. Traffic flow prediction is the task of predicting traffic volumes, utilizing
historical speed and volume data.
2. In TFT centralized machine learning methods use sufficient
sensor data, e.g. mobile, phone, camera etc. may contain users
private data. so there is privacy issue.
3. To solve privacy issue, privacy-preserving machine learning named
federated learning used.
4. In FL distributed organizations cooperatively train a globally
shared model through their local data without exchanging the raw data.
5. We apply Graph Convolutional (GCNN) to FL framework to
better capture the spatial-temporal dependency among traffic flow
data to improve predict accuracy while data privacy is preserve
3
NIT Delhi

Introduction
Fig: Privacy and security problems in traffic flow prediction. [1]

Motivation
1. Traffic flow prediction is the component of intelligent transportation
system(ITS) and can assist ITS to forecast and prevent traffic,
congestion control and manage traffic efficiently, and plan the best
travelling route.
2. One of the major motivations to use FL in TFT is to safeguard
data privacy.
3. Through the FL each organization is enable to train the model locally,
sharing only the model not private data.
4. GCNN give more accurate result as it capture dependency
among traffic data
5
NIT Delhi

Objective
1. Develop a accurate traffic prediction model while preserving privacy.
2. Improve accuracy of traffic flow prediction model.
3. Better use of sensitive data without sharing.
4. Understand how traffic flow data depend to time and space (location)
e.g. spatial- temporal dependency.
5. FL based GCNN model give better result with privacy preserve.
6
NIT Delhi

Problem Statement
1. A road network is considered as a directed graph G =(V, E, A),
with nodes V, represent the road links, edges E, denote road
intersection, A denotes the adjacency matrix of graph G.
2. We have to apply spatial and temporal graph convolution to
capture spatial-temporal dependency and secure parameter
aggregation mechanism to aggregate parameter at cloud.
3. According to secure parameter aggregation no traffic flow data is
exchange among detector stations.

Methodology
1. Federated Learning Framework: It consist following algorithms
FedAVG Algorithm: FedAVG consist following step
(i) The cloud select organizations to participate in this round
of training and broadcast global model to selected organization
(ii) Each organization train locally and updates the model
(iii) The cloud aggregates each organization’s model and update
global model
8
NIT Delhi

Methodology
Algorithm: FedAvg
Input : set of organization
output: w (weight )
1. initialize 𝑤𝑜 (pre-trained by a public dataset)
2. for each round do 3 to 8
3. {Ov} select participating organization
4. broadcast 𝑤𝑜 to organization in {Ov}
5. for each o in {Ov} do
6. initialize 𝑤𝑡
𝑜
= 𝑤𝑜
7. 𝑤𝑡+1
𝑜
Local Update(o, 𝑤𝑡
𝑜
)
8. 𝑤𝑡+1
𝑜
1/|Ov|* 𝑜 𝑖𝑛 𝑂𝑣 𝑤𝑡+1
𝑜
9. return(𝑤𝑡+1
𝑜
)
9
NIT Delhi

Methodology
2. Graph convolution in spatial dimension:
In spectral graph theory, a graph is represented by its Laplacian matrix:
L = D-A,
or its normalized form:
L = I-𝐷−1/2A𝐷1/2
Here A is the adjacency matrix of the graph, I is a unit matrix and D is
the
matrix with node degree values on the main diagonal:
Dii = 𝑗(𝐴𝑖𝑗)
10
NIT Delhi

Methodology
The eigenvalue decomposition of L can be written as:
L= UV𝑈𝑇
where V is diagonal matrix and, U is Fourier basis. The convolution
operation on graph is defined as the result of multiplying the signal x €
𝑅𝑛
on the graph with kernel g.
g*x = g(L)x = g(UV𝑈𝑇 )x = U g(V) 𝑈𝑇 x
Apply ReLU on g*x
11
NIT Delhi

Methodology
3. Graph convolution in temporal dimension:
For convolution in temporal dimension we take the result of
convolution in spatial dimension and multiply the parameters of the
convolutional kernel in temporal dimension.
Then apply ReLU on it as shows follow.
𝑋𝑝 = ReLU(h*(ReLU(g*𝑋𝑝−1)))
where g is kernel in spatial dimension, h is the parameters of the
convolutional
kernel in the temporal dimension, x is the original input time series or the
neural
result of calculation on the previous layer of the neural network.
12
NIT Delhi

Methodology
.
Fig: Federated learning framework [6]

Result
• Dataset Description:
• SZ-taxi: This data set have two matrix one for
road connectivity and another is for speed change
over time.
• Each matrix of data set have 156 row and 156
column, rows of matrix represent a particular
road and column represent corresponding values
of connectivity and traffic speed.

Contd…
.
Fig: shows the connectivity of different raods

Contd…
Fig: shows the speed change over time.

Contd...
• Evaluation Metrics:
• (1)Mean Absolute Error (MAE)
• (2) Root Mean squared Error(RMSE)
• (3) Accuracy
• 4) Variance Score

Contd…
Table: showing the result of different model.

Analysis
• We can see in previous table that r2 factor for HA model is negative it
means, HA model is not fit better for our graph data. RMSE and MAE
for HA model is more in compression to other model.
• In SVR model, error is reduce in compression to HA model. Hence
accuracy also improved.
• ARIMA model also have more error in compression to GCN model.
So its accuracy is also less.
• GCN and FedGCN have all the parameter nearly same. It because, in
FedGCN we use GCN model on local system them aggregate model
parameter on base station.
19
NIT Delhi

Contd…
• Result shows that if there is no spatial and temporal dependency then
we can use SVR, HA, ARIMA model.
• If dataset have spatial and temporal dependency but not privacy issue
then GCN is better.
• If our dataset have both spatial and temporal feature with privacy issue
then FedGCN is better.
20
NIT Delhi

Scope Of Work
1. Road safety aspect
2. Travel time prediction
3. It helps in automating roadways, railways etc.
4. It also helps in tracking and delivery of goods.
21
NIT Delhi

Conclusion and Future Work
• We conclude that for data having spatial and temporal features
GCN model work better.
• If there is spatial and temporal features in data with some
privacy issue then FedGCN is better
• In the future we will further optimize the network structure and
parameter settings.
• Moreover our proposed framework can be applied in to more
general spatio-temporal structured sequence forecasting
scenarios, such as evolving of social networks, and preference
prediction in recommendation system, etc.
22
NIT Delhi

References
[1] Y. Liu, J. J. Q. Yu, J. Kang, D. Niyato and S. Zhang, ”Privacy-Preserving
Traffic Flow Prediction: A Federated Learning Approach,” in IEEE Internet of
Things Journal, vol. 7, no. 8, pp. 7751-7763, Aug. 2020, doi:
10.1109/JIOT.2020.2991401.
[2] K. Chen et al., ”Dynamic Spatio-Temporal Graph-Based CNNs for Traffic
Flow Prediction,” in IEEE Access, vol. 8, pp. 185136-185145, 2020, doi:
10.1109/ACCESS.2020.3027375.
[3] A. Agafonov, ”Traffic Flow Prediction Using Graph Convolution Neural
Networks,” 2020 10th International Conference on Information Science and
Technology (ICIST), Bath, Lon- don, and Plymouth, United Kingdom, 2020, pp.
91-95, doi: 10.1109/ICIST49303.2020.9201971.
23
NIT Delhi

References
[4] Ya Zhang, Mingming Lu, Haifeng Li, ”Urban Traffic Flow Forecast Based on
FastGCR- NN” Journal of Advanced Transportation, vol. 2020, Article ID
8859538, 9 pages 2020. https://doi.org/1.1155/2020/8859538
[5] M. S. Ahmed, “Analysis of freeway traffic time series data and their
application to incident detection,” Equine Veterinary Education, vol. 6, no. 1, pp.
32–35, 1979. [6] Zhang, S., Tong, H., Xu, J., Maciejewski, R. (2018). Graph
Convolutional Networks: Algorithms, Applications and Open Challenges.
CSoNet.
[6] L. Zhao et al., "T-GCN: A Temporal Graph Convolutional Network for Traffic
Prediction," in IEEE Transactions on Intelligent Transportation Systems, vol. 21,
no. 9, pp. 3848-3858, Sept. 2020, doi : 10.1109/TITS.2019.2935152.
24
NIT Delhi

References
[7] Zhang, S., Tong, H., Xu, J., & Maciejewski, R. (2018). Graph Convolutional
Networks: Algorithms, Applications and Open Challenges. CSoNet.
[7] H. Huang, “Dynamic modeling of urban transportation networks and analysis
of its travel behaviors,” Chinese Journal of Management, vol. 2, pp. 18–22, Jan.
2005.
[8] Y. Zhao, M. Li, L. Lai, N. Suda, D. Civin, and V. Chandra, “Federated
learning with non-IID data,” 2018. [Online]. Available: arXiv:1806.00582.
[9] K. Bonawitz et al., “Towards federated learning at scale: System design,”
2019. [Online]. Available: arxiv.abs/1902.01046.
25
NIT Delhi

References
[10] J.Kang, Z. Xiong, D. Niyato, Y. Zou, Y. Zhang, and M. Guizani, “Reliable
federated learning for mobile networks,” 2019. [Online]. Available:
arXiv:1910.06837.
[11] C. Chao, “Freeway performance measurement system (PeMS),” Inst.
Transp. Stud., Univ. California at Berkeley, Berkeley, CA, USA, Rep. UCB-ITS-
PRR-2003-22, 2003.
[12] K. Xie et al., “An efficient privacy-preserving compressive data gathering
scheme in WSNs,” Inf. Sci., vol. 390, pp. 82–94, Jun. 2017
26
NIT Delhi

Federated learning based_trafiic_flow_prediction.ppt

More Related Content

What's hot (20)

Similar to Federated learning based_trafiic_flow_prediction.ppt (20)

Recently uploaded (20)

Federated learning based_trafiic_flow_prediction.ppt