Title Page
Contents
ABSTRACT 11
Ⅰ. Introduction 13
Ⅱ. Explaining Deep Learning Models for Tabular Data 16
2.1. Motivation 16
2.2. Related Work 20
2.2.1. Deep Learning Models 20
2.2.2. Explaining Deep Learning Model for Tabular Data with Time Causality 25
2.3. Problem Definition 28
2.4. Proposed Model 30
2.4.1. Identifying Time Lags 32
2.4.2. Building a Model Reflecting Time Lags 35
2.5. Performance Evaluation 37
2.5.1. Experiment Setup 37
2.5.2. Effectiveness Evaluation 40
Ⅲ. Explaining Deep Learning Models for Graph Data 45
3.1. Related Work 45
3.1.1. Deep Learning on Graphs 45
3.1.2. Interpretability of Deep Learning Models 48
3.2. Proposed Method 54
3.2.1. Overview 54
3.2.2. Node Classification using GCN 55
3.2.3. Explaining Node Classification Results 57
3.3. Performance Evaluation 60
3.3.1. Experiment Setup 60
3.3.2. Experimental Result on Synthetic Dataset 63
3.3.3. Experimental Result on Real Dataset 67
Ⅳ. Conclusion 70
REFERENCES 72
ABSTRACT IN KOREAN 77
Table 1. The influence of the important features on synthetic dataset found by the proposed method. 65
Table 2. Important features of the real dataset found by the proposed method. 67
Table 3. The influence of important features on real dataset found by the proposed method. 68
Figure 1. Examples of the time lag between explanatory and response variables. 18
Figure 2. The architecture of MLP. 20
Figure 3. The architecture of CNN. 21
Figure 4. The architecture of LSTM. 23
Figure 5. The form of the training dataset. 28
Figure 6. The proposed extension to an existing deep learning model. 31
Figure 7. An example of a transformed training dataset. 36
Figure 8. Performance evaluation results on the synthetic dataset. 42
Figure 9. Performance evaluation results on the real dataset. 44
Figure 10. The general design pipeline for a GNN model. 45
Figure 11. The architecture of GCN. 46
Figure 12. The architecture of Grad-CAM. 49
Figure 13. The architecture of LRP. 50
Figure 14. The intuition of LIME. 51
Figure 15. The concept of the proposed method. 52
Figure 16. The overall process of the proposed method. 54
Figure 17. Important features of the synthetic dataset found by the proposed method. 64