Title Page
Abstract
Contents
Chapter 1. Introduction 13
1.1. Background 13
1.2. Research Purpose and Scope 16
1.3. Research Contribution 18
Chapter 2. Literature Review 23
2.1. Traffic Speed and Congestion Prediction 23
2.2. Priority Link Identification 29
2.3. Link and Network Indices 35
2.4. Graph Attention Model 38
Chapter 3. Establishment of Impact on Adjacent Network Index 50
3.1. Index Setup 50
3.1.1. Research Flow and Data Description 50
3.1.2. Utilizing Speed Reduction Index 55
3.1.3. Creating an Impact on Adjacent Network Index 60
3.1.4. Preprocessing the Index 62
3.2. Analyzing the IANI 63
3.2.1. Statistical Property of IANI 63
3.2.2. Comparing IANI with Graph Centralities 74
3.2.3. Comparing IANII with SRI 76
Chapter 4. Graph Attention Model for Urban Network 83
4.1. Background of the Graph Attention Model 83
4.2. Improving the Graph Attention Model with Adjacency Matrix 89
4.2.1. The Traffic Flow Awareness Adjacency Matrix 89
4.2.2. Introducing Katz Centrality to the Adjacency Matrix 90
4.2.3. Handling the Overfitting and Oversmoothing Problem 93
4.3. Adding Physical Meaning to the Model 94
4.3.1. Reflecting the Heterogeneity of Road Networks 94
4.3.2. Incorporating Traffic Volume Data 98
4.3.3. Adding a Penalty as an Attention Loss 100
Chapter 5. Results 102
5.1. Improving the Adjacency Matrix 102
5.2. Considering Road Network Heterogeneity and Traffic Volume 114
5.3. The Result of Implementing Attention Loss and its Guidelines 118
Chapter 6. Conclusion 122
Bibliography 124
국문초록 130
Table 2.1. Literatures on Congestion Prediction 28
Table 2.2. Literatures on priority link identification 32
Table 2.3. Existing models and their evaluation with various criteria 35
Table 2.4. Definitions of each centrality 38
Table 2.5. Literature on traffic state prediction using graph attention model 42
Table 2.6. Indices of various networks 46
Table 3.1. Congestion Metrics and Assessment Criterias 57
Table 3.2. Statistics of a sum of future SRI 63
Table 3.3. Statistics about the number of N-hop links 65
Table 3.4. Statistics of IANI with SRI and normalized SRI 66
Table 3.5. Roads with the highest IANI mean value 68
Table 3.6. Roads with the highest IANI mean value at weekday and morning(6-10am) peak 69
Table 3.7. Roads with the highest IANI mean value at weekday and afternoon(5-9pm) peak 70
Table 3.8. Roads with the highest IANI mean value at weekend and morning(6-10am) peak 72
Table 3.9. Roads with the highest IANI mean value at weekend and evening(5-9pm) peak 73
Table 3.10. Correlation between various centralities and IANI 75
Table 3.11. Value of various centralities of the roads with the top 100 INAI 76
Table 3.12. Comparing the correlation of SRI and IANI with various centralities 77
Table 3.13. Various value of links by quintile of IANI 80
Table 3.14. Various value of links by quintile of SRI 81
Table 4.1. Example of an attention value matrix that is concentrated on a specific link 101
Table 5.1. Speed prediction result in Gangnam-gu 106
Table 5.2. Summary of Top 10 links in Gangnam-gu by summation of attention 111
Table 5.3. MAPE and its standard deviation for homogeneous and heterogeneous network 116
Table 5.4. MAPE and its standard deviation for without and with volume data 116
Table 5.5. Precision and recall error in the case of the model with and without traffic volume decoder 117
Table 5.6. Precision and recall error in the case of the model with and without attention loss 118
Table 5.7. Concentration loss and concentration loss count by the number of links 120
Figure 1.1. The purpose of the research 16
Figure 2.1. Failure case of a distance-based adjacency matrix 44
Figure 2.2. Degree distribution of Seoul road network(top) and conventional network (Airline) (bottom) 47
Figure 2.3. Speed of major links showing similar behavior 48
Figure 3.1. The framework and flow of the research 50
Figure 3.2. The service link of Seoul used in the research 55
Figure 3.3. Target service link of Gangnam-gu used in the research 56
Figure 3.4. Example of SRI for different link 58
Figure 3.5. SRI peaks with different values in the morning and afternoon 58
Figure 3.6. Discovery of unrelated links and delay in speed reduction propagation 59
Figure 3.7. Diagram of impact on adjacency matrix index 61
Figure 3.8. Before and after normalizing SRI 62
Figure 3.9. The Sum of future SRI before and after normalization 63
Figure 3.10. IANI with SRI and normalized SRI 66
Figure 3.11. Visualizing the IANI value 67
Figure 3.12. Mean values of various criteria by quintile of SRI and IANI 78
Figure 3.13. Mean values of various centralities by quintile of SRI and IANI 79
Figure 3.14. The different values between SRI and IANI in the same region 81
Figure 4.1. Example graph with five nodes and seven edges 84
Figure 4.2. Calculating the feature vector of node 2 in layer l 86
Figure 4.3. Temporal and spatial attention of ASTGCN 87
Figure 4.4. Iterative scheme of Spatio-temporal attention block 88
Figure 4.5. The traffic flow awareness adjacency matrix 89
Figure 4.6. k-times multiplying adjacency matrix enables to connect k-hop matrix 91
Figure 4.7. The example of oversmoothing caused by extensive receptive field 93
Figure 4.8. Propagation of speed reduction through an uninterrupted flow. Each subfigure's time step starts from 6:00 am to 8:00 30 min interval 95
Figure 4.9. Propagation of speed reduction through an interrupted flow. Each subfigure's time step starts from 6:00 am to 8:00 30 min interval 96
Figure 4.10. Visualization of uninterrupted flow(red) and interrupted flow(light blue) 97
Figure 4.11. The structure of Attention-based Spatio-Temporal Heterogeneous Graph Convolution Network (AST-HGCN) 98
Figure 4.12. Data points with traffic volume 99
Figure 4.13. Incorporating traffic volume using a decoder neural network 99
Figure 5.1. Input and prediction horizon and scheme 103
Figure 5.2. Speed prediction result of future 5min(top) and 60min(bottom) of Gangnam-gu link (selected) 104
Figure 5.3. Prediction Horizon by RMSE loss of Gangnam-gu links 105
Figure 5.4. The illustration of the term "attention value" 107
Figure 5.5. Attention Sum Histogram of Links in Gangnam-gu 108
Figure 5.6. Top 10 links in Gangnam-gu by summation of attention 110
Figure 5.7. Attention value of top 3 links of Gangnam-gu on Tuesday, November 2, 2021, afternoon 112
Figure 5.8. Attention value of top 3 links of Gangnam-gu on Wednesday, November 3, 2021, around the lunchtime 113
Figure 5.9. Example to introduce precision and recall (Let the true is trying to judge the top-3 samples) 115
Figure 5.10. Concentration loss ratio by number of links 121