Title Page
ABSTRACT
국문 초록
Contents
CHAPTER 1. INTRODUCTION 13
CHAPTER 2. RELATED WORKS 16
CHAPTER 3. PROBLEM FORMULATION 19
CHAPTER 4. PROPOSED MODEL 21
4.1. Multi-Scale Dilated Convolution Module 23
4.2. Input-Adaptive Graph Construction Module 24
4.3. Graph Convolution Module 28
4.4. Scale-Wise Propagation Module 31
CHAPTER 5. EXPERIMENTAL SETTINGS 33
5.1. Datasets 33
5.2. Experimental Settings 35
5.3. Evaluation Metrics 36
5.4. Comparison Methods 37
CHAPTER 6. RESULTS AND DISCUSSION 38
6.1. Multivariate Time-Series Forecasting 38
6.2. Ablation Study 45
6.3. Parameter Study 47
6.4. Graph Analysis 49
CHAPTER 7. CONCLUSION 54
REFERENCES 55
Table 1. Summary of Datasets 34
Table 2. Comparison of the Forecasting Performance for Solar-Energy Dataset 39
Table 3. Comparison of the Forecasting Performance for Traffic Dataset 40
Table 4. Comparison of the Forecasting Performance for Electricity Dataset 41
Table 5. Comparison of the Forecasting Performance for Exchange-Rate Dataset 42
Table 6. Ablation Study Results for Solar-Energy Dataset 46
Table 7. Ablation Study Results for Electricity Dataset 46
Figure 1. Problem Formulation 20
Figure 2. The Full Framework of the Proposed Model 22
Figure 3. Input-Adaptive Graph Construction Module 26
Figure 4. Graph Convolution Module 30
Figure 5. Autocorrelation Graph of Solar-Energy Dataset 43
Figure 6. Autocorrelation Graph of Traffic Dataset 43
Figure 7. Autocorrelation Graph of Electricity Dataset 44
Figure 8. Autocorrelation Graph of Exchange-Rate Dataset 44
Figure 9. RRSE Result in Different Number of Scales for Solar-Energy Dataset 48
Figure 10. RRSE Result in Different Number of Scales for Electricity Dataset 48
Figure 11. Example of Obtained Dynamic Adjacency Matrices (Solar-Energy) 50
Figure 12. Example of Obtained Dynamic Adjacency Matrices (Traffic) 51
Figure 13. Example of Obtained Dynamic Adjacency Matrices (Electricity) 52
Figure 14. Example of Obtained Dynamic Adjacency Matrices (Exchange-Rate) 53