Meta-Learning for Graphs with Heterogeneous Node Attribute Spaces for Few-Shot Edge Predictions
Abstract: Prediction of edges between nodes in graph data is useful for many applications, such as social network analysis and knowledge graph completion. Existing graph neural network-based approaches have achieved notable advancements, but encounter significant difficulty in building an effective model when there is an insufficient number of known edges in graphs. Although some meta-learning approaches were introduced to solve this problem, having an assumption that the nodes of training graphs and test graphs are in homogeneous attribute spaces, which limits the flexibility of applications. In this paper, we proposed a meta-learning method for edge prediction that can learn from graphs with nodes in heterogeneous attribute spaces. The proposed model consists of attribute-wise message-passing networks that transform information between connected nodes for each attribute, resulting in attribute-specific node embeddings. The node embeddings are obtained by calculating the mean of the attribute-specific node embeddings.The encoding operation can be repeated multiple times to capture complex patterns. The attribute-wise message-passing networks are shared across all graphs, allowing knowledge transfer between different graphs.The probabilities of edges are estimated by the Euclidian distance between node embeddings. Experimental results on 14 real-world data sets demonstrate that the proposed method outperforms existing methods in edge prediction problems with sparse edge information.
Submission Length: Regular submission (no more than 12 pages of main content)
Changes Since Last Submission: Dear Action Editor,
Thank you again for your constructive feedback on our paper.
To ensure fairness in our comparisons, we have adjusted the model complexity of all comparative models to a similar level and aligned the training time. For DSGCN and MAMLGCN, we increased the number of hidden units in each feed-forward neural network from 32 to 64 and raised MAMLGCN's training epochs from 100 to 500.
For the non-meta-learning approaches (NN, GCN, and GAT), we increased the number of neural network layers from three to five and the hidden units from 32 to 64. The updated results are presented in Table 3.
Even after aligning model complexity and training time across the comparative models, HGML still achieves the best performance.
In addition to revising the experiments on the comparative models, we have updated Table 4 to include an analysis of the number of parameters, average training time, and inference time on new input graphs. We have also revised the description in Section 5.1 accordingly.
> Table 4 presents the number of parameters, average training time for meta-learning, and inference time. The inference time refers to the time required for the model to process a query graph and generate predictions for the query edges, including the local adaptations of MAMLGCN and the re-training required for NN, GCN, and GAT. To evaluate inference time, we sample three different graph sizes: ( N = 500, 5,000, 20,000) from Reddit, ensuring that each sampled graph contains N nodes and N edges to simulate the few-shot scenario. HGML achieves better performance while maintaining similar model complexity and training time compared to DSGCN and MAMLGCN. MAMLGCN and the non-meta-learning approaches have higher inference time. In contrast, HGML can be applied to new graphs with a reasonable inference time.
Assigned Action Editor: ~Olgica_Milenkovic1
Submission Number: 3400
Loading