The document proposes a Deep Incident-Aware Graph Convolutional Network (DIGC-Net) model to predict traffic speeds based on incidents. DIGC-Net uses a graph convolutional network to capture spatial road features, an LSTM to capture temporal patterns, and a fully connected layer to capture periodic features. It also uses an RNN to model how past incidents affect future traffic and speeds. The model discovers critical incidents using anomalous degree and speed variation, extracts latent incident impact features using a classifier, and incorporates these along with spatio-temporal and periodic features for traffic speed prediction.

1. Deep Graph Convolutional Networks for
Incident-Driven Traffic Speed Prediction
ACM Conference on Information and Knowledge Management, 2020
Qinge Xie, Tiancheng Guo, Yang Chen, Yu Xiao, Xin Wang, and Ben Y. Zhao
March 12, 2021
Presenter: KyungHwan Moon

2. Contents
• Overview of the Paper
• Background and Motivation
• Introduction
• Deep Incident-Aware Graph Convolutional Network
• Experiment
• Conclusion and Discussion

3. 2
Overview of the Paper
Spatio-temporal Learning
(Graph Convolutional Network + LSTM)
 Deep Incident-Aware Graph Convolutional Network
∎ Graph Convolutional Network (GCN)
: Capture spatial features of road networks
∎ LSTM
: Capture the time evolution patterns of traffic speeds
∎ RNN
: Contains loops that allow information to persist, so
previous incidents will affect the traffic conditions,
which may lead to the occurrence of future incidents
∎ Fully Connected Layer (FC)
: Capture the long-term periodic features
Incident Learning
(RNN)
Periodic Learning
(Fully Connected Layer)

4. 3
Background and Motivation
 Previous studies on traffic speed prediction predominately used
spatio-temporal and context features for prediction
∎ They have not made good use of the impact of traffic incidents
 Incident-driven prediction framework consists of three processes
∎ Propose a critical incident discovery method to discover traffic incidents with
high impact on traffic speed
∎ Design a binary classifier, which uses deep learning methods to extract the
latent incident impact features
∎ Propose DIGC-Net to effectively incorporate traffic incident, spatio-temporal,
periodic and context features for traffic speed prediction

5. 4
Introduction
 Traffic speed prediction has been a challenging problem for decades
∎ Congestion control [17]
∎ Vehicle routing planning [14]
∎ Urban road planning [28]
∎ Travel time estimation [9]
 There are two main challenges for incident-driven traffic speed prediction problem
∎ The impact of traffic incidents is complex and varies significantly across incidents
- It is unreasonable to treat all traffic incidents equally for traffic speed prediction,
which may even negatively impact the prediction performance
∎ The impact of traffic incidents on adjacent roads will be affected by external factors
like incident occurrence time, incident type and the road topology structure
- Need to extract the latent impact features of traffic incidents to improve
the traffic prediction

6. 5
Introduction
 Propose a critical incident discovery method to quantify the impact of urban traffic
incidents on traffic flows to tackle the first challenge
∎ Consider both anomalous degree and speed variation of adjacent roads to discover the
critical traffic incidents
 Propose a binary classifier which uses deep learning methods to extract the latent
impact features of incidents to tackle the second challenge
∎ Extract the latent impact features from the middle layer of the classifier, where the latent
features are continuous and filtered

7. 6
Introduction

8. 7
Introduction
 Datasets
∎ San Francisco (SFO)
∎ New York City (NYC)
 Problem Formulation and Preprocessing
∎ Reconstruction of the road network
- Use the road segment as the node, and use every flow as one node to build the road
network more specifically
- If two flows have points of intersection, add an edge to connect node and node
∎ Problem formulation
- Use to represent the speed of flow at time slot t
- For every speed snapshot of the road network, get a vector of all flows
(N is the total number of flows)
- Given the re-build road graph and a T-length historical real-time speed sequence of all flows,
task is to predict future speeds of every flow in the city where k is the prediction length

9. 8
Deep Incident-Aware Graph Convolutional Network
(DIGC-Net)
 Methodology
∎ Congestion Incident
- M : center point of the incident
- r : the radius of the impact range
- The circle with the center M and
radius r stands for the region
affected by the incident
- Define that if the center of flows is in the circle, then
the flows might be affected by the incident
- The blue, red and green lines represent three flows which
might be affected by the incident, respectively
Urban Critical Incident Discovery
Extract The Latent Incident
Impact Features
Incident-driven
traffic speed prediction

10. 9
Deep Incident-Aware Graph Convolutional Network
(DIGC-Net)
 Methodology
∎ Congestion Incident
- Analyze each candidate flow that whether it will truly be
affected by the incident
- Use a variant of the method proposed in [41] to compute
the anomalous degree of each flow
- To compute the anomalous degree of a region is based on
its historically similar regions in the city
- The sudden drop of speed similarity of a region and its
historically similar regions indicates the occurrence of
urban anomalies
Urban Critical Incident Discovery
Extract The Latent Incident
Impact Features
Incident-driven
traffic speed prediction

11. 10
Deep Incident-Aware Graph Convolutional Network
(DIGC-Net)
 Methodology
∎ Pair-wise Similarity of Flows
- The pair-wise similarity is calculated by
: 𝑠𝜉𝑖,𝜉𝑗
[𝑡−𝑇+1:𝑡]
= 𝑃(𝑣𝜉𝑖
𝑡−𝑇+1:𝑡
, 𝑣𝜉𝑗
[𝑡−𝑇+1:𝑡]
)
where P is to calculate Pearson correlation coefficient
[20] of two speed sequences
- The similarity matrix S of all flows at t is calculated by
the following equation:
𝑆𝑡
=
𝑠𝜉0,𝜉0
[𝑡−𝑇+1:𝑡]
⋯ 𝑠𝜉0,𝜉𝑁−1
[𝑡−𝑇+1:𝑡]
⋮ ⋱ ⋮
𝑠𝜉𝑁−1,𝜉0
[𝑡−𝑇+1:𝑡]
⋯ 𝑠𝜉𝑁−1,𝜉𝑁−1
[𝑡−𝑇+1:𝑡]
where N is the total number of flows in the city
Urban Critical Incident Discovery
Extract The Latent Incident
Impact Features
Incident-driven
traffic speed prediction

12. 11
Deep Incident-Aware Graph Convolutional Network
(DIGC-Net)
 Methodology
∎ Similarity Decrease Matrix D
- The decreased similarity of each flow pair from time slot
t-1 to t
- D at time slot t is calculated by:
𝐷𝑡 = max(0, 𝑆𝑡−1 − 𝑠𝑡)
∎ Anomalous Degree A
- Use similarity matrix S and similarity decrease matrix D
- Use a threshold parameter 𝛿 to capture the historically similar flows
- When the similarity of two flows is greater than or equal
to 𝛿, define that they are historically similar
Urban Critical Incident Discovery
Extract The Latent Incident
Impact Features
Incident-driven
traffic speed prediction

13. 12
Deep Incident-Aware Graph Convolutional Network
(DIGC-Net)
 Methodology
∎ Anomalous Degree A
- Given a flow 𝜉𝑖 at time slot t, the historically similar flow
sets of 𝜉𝑖 is denoted as
𝐻𝜉𝑖
𝑡
= 𝜉𝑗 𝑖 ≠ 𝑗 𝑎𝑛𝑑 𝑆𝑖,𝑗
𝑡
= 𝑆𝑗,𝑖
𝑡
≥ 𝛿}
- Anomalous degree of flow 𝜉𝑖 at time slot t is calculated by
the following equation:
𝐴𝜉𝑖
𝑡
=
𝜉𝑗∈𝐻𝜉𝑖
𝑡 𝑆𝑖,𝑗
𝑡−1
∙ 𝐷𝑖,𝑗
𝑡
𝜉𝑗∈𝐻𝜉𝑖
𝑡 𝑆𝑖,𝑗
𝑡−1
where A is the decrease degree in speed similarity of 𝜉𝑖
and its historically similar flows
Urban Critical Incident Discovery
Extract The Latent Incident
Impact Features
Incident-driven
traffic speed prediction

14. 13
Deep Incident-Aware Graph Convolutional Network
(DIGC-Net)
 Methodology
∎ Anomalous Degree A
Urban Critical Incident Discovery
Extract The Latent Incident
Impact Features
Incident-driven
traffic speed prediction

15. 14
Deep Incident-Aware Graph Convolutional Network
(DIGC-Net)
 Methodology
∎ Local Anomalous Degree Algorithm
- It will cost a lot to compute the similarity matrix S, the
similarity decrease matrix D and the anomalous degree A
- Propose a local anomalous degree algorithm to speed up
our method based on the spectral clustering algorithm [39]
- Assume that traffic in nearby locations should be similar[32, 36, 45]
- Assume that flows in the same community and in the
spatially nearby regions will be historically similar
- Given, a graph G, perform spectral decomposition and
obtain k graph spatial features of each flow
- Use K-means [8], a common unsupervised clustering
method, to cluster flows into k classes
Urban Critical Incident Discovery
Extract The Latent Incident
Impact Features
Incident-driven
traffic speed prediction

16. 15
Deep Incident-Aware Graph Convolutional Network
(DIGC-Net)
 Methodology
∎ Local Anomalous Degree Algorithm
Urban Critical Incident Discovery
Extract The Latent Incident
Impact Features
Incident-driven
traffic speed prediction

17. 16
Deep Incident-Aware Graph Convolutional Network
(DIGC-Net)
 Methodology
∎ Validation of Local Algorithm
- Eigenvectors can effectively capture spatial graph features
- Only need to compute the local values of the similarity matrix 𝑆,
the similarity decrease matrix 𝐷 and the anomalous degree 𝐴
in the same district
- Explore the impact on traffic flows of different urban traffic incidents
Urban Critical Incident Discovery
Extract The Latent Incident
Impact Features
Incident-driven
traffic speed prediction

18. 17
Deep Incident-Aware Graph Convolutional Network
(DIGC-Net)
 Methodology
∎ Validation of Local Algorithm
Urban Critical Incident Discovery
Extract The Latent Incident
Impact Features
Incident-driven
traffic speed prediction

19. 18
Deep Incident-Aware Graph Convolutional Network
(DIGC-Net)
 Methodology
∎ Relative Speed Variation 𝑅
- Given a flow 𝜉𝑖 at time t, and the historical speed sequence
[𝑣𝜉𝑖
𝑡−𝑇+1
, 𝑣𝜉𝑖
𝑡−𝑇+2
, ⋯ , 𝑣𝜉𝑖
𝑡
] of 𝜉𝑖 in a T-length time window
- Define the relative speed variation of 𝜉𝑖 as follow:
𝑅𝜉𝑖
𝑡
=
𝑡′=𝑡−𝑇+1
𝑡′=𝑡
𝑣𝜉𝑖
𝑡′
𝑇
− 𝑣𝜉𝑖
𝑡
max(𝑣𝜉𝑖
𝑡𝑠
, 𝑣𝜉𝑖
𝑡𝑠+1
, ⋯ , 𝑣𝜉𝑖
𝑡𝑒
)
use 24 hours (288 intervals) as the normalization window length,
i.e., 𝑡𝑠 = 𝑡 − 144 𝑎𝑛𝑑 𝑡𝑒 = 𝑡 + 144, 𝑎𝑛𝑑 𝑇 = 10 𝑖𝑛𝑡𝑒𝑟𝑣𝑎𝑙𝑠
Urban Critical Incident Discovery
Extract The Latent Incident
Impact Features
Incident-driven
traffic speed prediction

20. 19
Deep Incident-Aware Graph Convolutional Network
(DIGC-Net)
 Methodology
∎ Relative Speed Variation 𝑅
Urban Critical Incident Discovery
Extract The Latent Incident
Impact Features
Incident-driven
traffic speed prediction

21. 20
Deep Incident-Aware Graph Convolutional Network
(DIGC-Net)
 Methodology
∎ Validation of Relative Speed Variation
- Consider three related features:
- Slope of speed variation 𝑘 [33]
- Recent speed 𝑣𝑡−1 [2]
- Historical average speed (𝑣) [2]
corresponding to three candidate computing methods of
Relative Speed Variation 𝑅
1) 𝑅𝑘+𝑣𝑡−1+𝑣 = 𝑣 − 𝑣𝑡 × 𝑘 × 𝑝 + 𝑣𝑡−1 − 𝑣𝑡 × 𝑘𝑡−1 × 𝑞
2) 𝑅𝑣𝑡−1+𝑣 = 𝑣 − 𝑣𝑡
× 𝑝 + 𝑣𝑡−1
− 𝑣𝑡
× 𝑞
3) 𝑅𝑣 = |𝑣 − 𝑣𝑡|
where p and q set to 0.5
Urban Critical Incident Discovery
Extract The Latent Incident
Impact Features
Incident-driven
traffic speed prediction

22. 21
Deep Incident-Aware Graph Convolutional Network
(DIGC-Net)
 Methodology
∎ Incident Effect Score 𝜀
- Due to the complementarity of anomalous degree and relative
speed variation, combine both of them to compute the incident
effect score
Given a flow 𝜉𝑖 at time slot t, the incident effect score is calculated by:
𝜀𝜉𝑖
𝑡
= 𝜌 ∙ 𝐴𝜉𝑖
𝑡
+ (1 − 𝜌) ∙ 𝑅𝜉𝑖
𝑡
where 𝜌 is a parameter to control the ratio of 𝐴 and 𝑅
Urban Critical Incident Discovery
Extract The Latent Incident
Impact Features
Incident-driven
traffic speed prediction

23. 22
Deep Incident-Aware Graph Convolutional Network
(DIGC-Net)
 Methodology
∎ Critical Incidents
- For incidents like mega-events, the traffic flows might be affected
before incidents begin. On the contrary, incidents like traffic collisions
will begin to affect traffic flows after they occurred
Define the flows which are highly affected by accident as
𝜉𝑖 max(𝜀𝜉𝑖
𝑡−
𝑇
2
, 𝜀𝜉𝑖
𝑡−
𝑇
2+1
, ⋯ , 𝜀𝜉𝑖
𝑡+
𝑇
2
) ≥ 𝜃}
where 𝜃 is a threshold parameter
When | 𝜉𝑖 max 𝜀𝜉𝑖
𝑡−
𝑇
2
, 𝜀𝜉𝑖
𝑡−
𝑇
2
+1
, ⋯ , 𝜀𝜉𝑖
𝑡+
𝑇
2
≥ 𝜃 |𝐼𝑘
> 0
there is at least one flow is highly affected by the incident 𝐼𝑘, we call
𝐼𝑘 is a critical incident
Urban Critical Incident Discovery
Extract The Latent Incident
Impact Features
Incident-driven
traffic speed prediction

24. 23
Deep Incident-Aware Graph Convolutional Network
(DIGC-Net)
 Methodology
Urban Critical Incident Discovery
Extract The Latent Incident
Impact Features
Incident-driven
traffic speed prediction

25. 24
Deep Incident-Aware Graph Convolutional Network
(DIGC-Net)
 Methodology
∎ Spatial Learning : GCN
- Adapt graph convolutional network (GCN) to learn the spatial
topology features.
(To capture the topology features in non-Euclidean structures, which
is suitable for road networks)
∎ Temporal Learning : LSTM
- Adapt Long Short-Term Memory (LSTM) model as temporal learning
component
(To learn long-term dependency information of time related
sequences)
Urban Critical Incident Discovery
Extract The Latent Incident
Impact Features
Incident-driven
traffic speed prediction

26. 25
Deep Incident-Aware Graph Convolutional Network
(DIGC-Net)
 Methodology
∎ Context Learning
- Use the following features for context learning
1) Incident type
(Traffic collision and event)
2) Road status
(An incident leads to a road close or not)
3) Start and end hour
(Start time 𝑡𝑠 and an anticipative end time 𝑡𝑒 of an incident)
4) Incident duration
(The anticipative duration of an incident
5) Weekday, Saturday or Sunday
Urban Critical Incident Discovery
Extract The Latent Incident
Impact Features
Incident-driven
traffic speed prediction

27. 26
Deep Incident-Aware Graph Convolutional Network
(DIGC-Net)
 Methodology
∎ Latent incident impact features extraction
Urban Critical Incident Discovery
Extract The Latent Incident
Impact Features
Incident-driven
traffic speed prediction

28. 27
Deep Incident-Aware Graph Convolutional Network
(DIGC-Net)
 Methodology
∎ Spatio-temporal Learning
- Use the similar structure of spatial and temporal learning of binary
classifier. The output of spatio-temporal learning is 𝑌𝑠
∎ Incident Learning
- Select all incidents occurred within [t-125min, t-5min] as the incident
learning inputs (the last two hours) and use the pre-trained binary
classifier to extract (𝑌𝑐 ⊕ 𝑌
𝑔)𝐹𝐶𝑠, i.e., the latent incident impact
features of each incident
∎ Periodic Learning
- Use the same time slots in the last 5 days to learn the periodic
features. The output of periodic learning 𝑌𝑃
Urban Critical Incident Discovery
Extract The Latent Incident
Impact Features
Incident-driven
traffic speed prediction

29. 28
Deep Incident-Aware Graph Convolutional Network
(DIGC-Net)
 Methodology
∎ Spatio-temporal Learning
- Use the similar structure of spatial and temporal learning of binary
classifier
∎ Incident Learning
- Select all incidents occurred within [t-125min, t-5min] as the incident
learning inputs (the last two hours)
∎ Periodic Learning
- Use the same time slots in the last 5 days to learn the periodic
features
Urban Critical Incident Discovery
Extract The Latent Incident
Impact Features
Incident-driven
traffic speed prediction

30. 29
Experiment
Urban Critical Incident Discovery
Extract The Latent Incident
Impact Features
Incident-driven
traffic speed prediction
 Results

31. 30
Experiment
Urban Critical Incident Discovery
Extract The Latent Incident
Impact Features
Incident-driven
traffic speed prediction
 Results

32. 31
Experiment
Urban Critical Incident Discovery
Extract The Latent Incident
Impact Features
Incident-driven
traffic speed prediction
 Results

33. 32
Experiment
Urban Critical Incident Discovery
Extract The Latent Incident
Impact Features
Incident-driven
traffic speed prediction
 Results

34. 33
Experiment
 Conclusion and Discussion
∎ Propose a critical incident discovery method to identify urban crucial
incidents and their impact on traffic flows
∎ Design a binary classifier to extract the latent incident impact features for
improving traffic speed prediction
∎ Propose a Deep Incident-Aware Graph Convolutional Network (DIGC-Net)
to effectively incorporate traffic incident, spatio-temporal, periodical and
weather features for traffic speed prediction

35. Thank you
Any questions?