Loading...

Transfer Learning in Graph Analysis

Mastani, Mahdi | 2025

0 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 58351 (02)
  4. University: Sharif University of Technology
  5. Department: Mathematical Sciences
  6. Advisor(s): Razvan, Mohammad Reza; Rabiei. Hamid Reza
  7. Abstract:
  8. Graph-structured data plays a pivotal role in real-world applications. However, graph-based machine learning models, such as Graph Neural Networks (GNNs), often face the challenge of domain shift. This phenomenon, stemming from the statistical distribution mismatch between the source and target domains, limits the generalizability of models, especially in graph settings where shifts occur simultaneously in both node attributes and network structure. This thesis addresses the problem of Graph Domain Adaptation (GDA) and proposes a novel framework named "Two-Stage Graph Domain Adaptation" (TGDA). The primary objective of TGDA is to effectively and disentangle the two main sources of domain shift: "node attribute shift" and "conditional structure shift." In its first stage, TGDA performs a structure-free alignment of node attributes using domain adversarial learning to produce representations that are both domain-invariant and discriminative. In the second stage, these aligned representations are fed, along with the graph structure of each domain, into a GNN, and the resulting structure-aware representations are then re-aligned using another domain adaptation mechanism. The performance of TGDA was evaluated on five real-world datasets and one synthetic dataset. The results consistently demonstrate the superiority of TGDA over baseline GDA methods, particularly in the presence of simultaneous attribute and structure shifts. Ablation studies confirmed the importance of both adaptation stages and the role of adversarial training. Furthermore, the flexibility of TGDA with respect to different GNN architectures and the relative superiority of DANN-based alignment over MMD (especially in terms of scalability) were observed.
  9. Keywords:
  10. Domain Adaptation ; Graph Neural Network ; Adversarial Training ; Domain Shift ; Graph Domain Adaptation

 Digital Object List

 Bookmark

...see more