High-order Graph Neural Networks with Common Neighbor Awareness for Link Prediction
The study presented in this paper tackles an underexplored aspect of dynamic graph learning (DGL)-known as link prediction-using high-order graph neural networks (HGNN). The focus is on enhancing the predictive accuracy of dynamic graph neural networks (DGNNs) by incorporating common neighbor awareness (CNA) into the learning process. The paper introduces the High-order Graph Neural Network with Common Neighbor Awareness (HGNN-CNA), designed to address significant limitations in existing approaches that primarily consider only pairwise node interactions.
Problem Definition and Context
In DGL, link prediction is a critical task, especially relevant in areas like recommendation systems and biological networks. Traditional DGNNs rely extensively on a message passing framework that aggregates information based on direct node interactions. However, such models often overlook multifaceted interactions arising from common neighbors. The research asserts that common neighbor interactions play a substantial role in graph structure and therefore ought to be integrated into the model.
HGNN-CNA Approach
The HGNN-CNA model brings two innovative aspects to the table:
Common Neighbor Correlation: By leveraging multi-hop common neighbor information, HGNN-CNA can capture intricate interaction patterns among nodes. This is achieved by estimating a correlation score that acknowledges not only direct, but also indirect interactions mediated through common neighbors.
Message Passing Integration: The model integrates this common neighbor correlation score directly into the message passing scheme. By doing so, it redefines aggregation weights to include interactions of nodes that share common neighbors, thereby enriching node embeddings and improving link prediction accuracy.
Methodology
The methodology involves several key components:
Structural Feature Learning: The model generates node-level structural features from the adjacency matrix of the DG, facilitating the capture of common neighbors.
Correlation Score Calculation: By employing the tensor product, the model captures spatial-temporal patterns, which contribute to obtaining a refined aggregation weight for the message passing process.
Model Optimization: The learning objective is formulated in a way that combines link prediction probability with regularization terms to optimize the DG representation.
Results
Experimental validation is conducted using three real-world DG datasets, indicating that HGNN-CNA outperforms multiple state-of-the-art DGNN-based models. Notably, it achieves higher F1-scores and accuracy across various datasets, showcasing a significant performance boost in link prediction tasks. For example, in the dataset D1, HGNN-CNA achieved an F1-Score of 0.8496, which is distinctly superior to competing models.
Implications and Future Directions
The introduction of common neighbor awareness in link prediction suggests broader implications. Firstly, the methodology could be generalized to include higher-order interactions in other graph-based tasks, potentially improving the quality of embeddings beyond link prediction. Secondly, integrating common neighbor correlations with message-passing architectures may illustrate new pathways to fuse graph structure information in deep learning models.
Looking forward, the study opens possibilities for further research into varying message-passing schemes that incorporate these correlations differently, potentially enhancing robustness, scalability, and efficiency. There's also room to investigate the impact of varying the number of hops considered in common neighbor correlations, which may influence the depth of insights into graph structures.
In conclusion, HGNN-CNA provides a substantial advancement in link prediction on dynamic graphs, emphasizing common neighbors’ roles and proposing a powerful paradigm shift from traditional DGNNs to more comprehensive frameworks.