Title Page
Abstract
Contents
Chapter 1. Introduction 8
Chapter 2. Related Work 11
2.1. Graph-based PU learning 11
2.2. Mixup for graph data 12
Chapter 3. Problem Definition 13
Chapter 4. Methodology 15
4.1. GNN 16
4.2. Mixup strategy 16
4.2.1. Mixup between XP and XU.[이미지참조] 17
4.2.2. Mixup between XP (XU) and XP (XU).[이미지참조] 18
4.3. Optimization 19
4.3.1. Original PU optimization 19
4.3.2. PUM-GNN optimization 20
Chapter 5. Experiment 22
5.1. Dataset 22
5.2. Baselines 23
5.2.1. positive-unlabeled 23
5.2.2. positive-negative 24
5.3. Implementation details 25
5.4. Experimental results 26
5.4.1. Node classification performance (RQ1) 26
5.4.2. Effect of πp(RQ2)[이미지참조] 32
5.4.3. Analysis of negative-prediction preference (RQ3) 34
5.4.4. Analysis of training session (RQ4) 36
5.4.5. Parameter Sensitivity (RQ5) 39
Chapter 6. Conclusion 42
Bibliography 43
초록 48
Table 5.1. Dataset 23
Table 5.2. Node classification performance under PU labels, showing average F1 score. 27
Table 5.3. Node classification performance under PN labels, showing average F1 score. 30
Table 5.4. Node classification performance under reversed πp, showing average F1 score.[이미지참조] 33
Table 5.5. Node classification performance under different α values. 41
Figure 3.1. Positive-Unlabeled classification with πp.[이미지참조] 14
Figure 4.1. Overview of PUM-GNN. 15
Figure 4.2. Mixup strategy of PUM-GNN. 17
Figure 5.1. Analysis of false negatives and true positives. 35
Figure 5.2. The training curves of PU-GCN and PUM-GCN.(up: Cora, down: Citeseer) 37
Figure 5.3. t-SNE visualization of trained embedding.(up: PU-GCN, down: PUM-GCN) 38
Figure 5.4. Parameter analysis with respect to # of layers. 40