Decoding the Neighborhood Aggregation Mechanism of GNNs: From Generalization to Interpretive Analysis
- Publication Type:
- Thesis
- Issue Date:
- 2023
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
Graph Neural Networks (GNNs) have emerged as a powerful tool for analyzing graph-structured data, enabling effective graph representation learning and capturing complex dependencies among nodes. This thesis presents three research works that contribute to the advancement of GNNs in generalization, scalability, and interpretability.
The first research work proposes a general and scalable GNN framework called Virtual Adjacency Graph Neural Network (VAGNN). VAGNN introduces the concept of an optimizable virtual adjacency matrix, enabling the flexible construction of neighborhood sets for aggregation in GNNs. It overcomes the limitations of existing GNNs by customizing the inclusion of local and global nodes, offering a selection of attention mechanisms for the aggregation function, and providing the ability to incorporate supplementary information for further control over the aggregation process.
The second work focuses on the interpretive analysis of GNNs in link prediction. By exploring the underlying mechanisms of GNNs in link prediction, this research investigates the learning of pair-specific structural information and the effectiveness of node embeddings in GNNs. This work reveals the limitations of existing GNN-based link prediction models and highlights the importance of incorporating common neighbor-based heuristics. Empirical evaluations on real-world datasets support our findings and provide insights for the design of more robust link prediction algorithms.
The third research work undertakes an investigation into the inner working mechanism of diverse GNN techniques in real-world scenarios. Our objective revolves around identifying the challenges that arise in practical problems, proposing subsequent solutions to these challenges through the utilization of various techniques, and providing insightful interpretations regarding the effectiveness of these techniques. We introduce a GNN-based framework called Multi-hop Attention-based GNN (MAG). This framework develops a novel geographical graph construction method, aiming to capture the geographical neighboring effects among nodes. Furthermore, MAG addresses a multitude of challenges, including differentiating information from different neighbors through an attention mechanism, mitigating the over-smoothing issue via the use of residual connections, and capturing temporal effects with the integration of a point process module.
These research works contribute to the development and understanding of GNNs. They present novel GNN frameworks, provide insights into the strengths and limitations of GNNs in link prediction, and interpret the working mechanism of different techniques in GNNs. The findings and methodologies presented in this thesis contribute to advancing the field of graph representation learning and open up new possibilities for analyzing complex systems represented as graphs.
Please use this identifier to cite or link to this item: