SmGNN: Link Prediction in Sparse Layers of Multi-layer Graphs
Abstract: Link prediction is a crucial task in multi-layer graphs for different applications, where real-world graphs often consist of multiple types of relations represented as different layers. However, these multi-layer graphs often suffer from missing edges, especially in specific layers with a high number of missing edges (sparse layers) due to privacy concerns. In this paper, we tackle the challenge of predicting missing links in such layers to enhance the link prediction performance in multi-layer graphs. Training a Graph Neural Network (GNN) directly for link prediction on the sparse layer with limited edges would be challenging for exploring missing links and may lead to sub-optimal performance. To tackle this problem, we propose a novel framework called Sparse Layer Reconstruction Multi-layer Graph Neural Network (SmGNN). SmGNN proposes to leverage information from other relation types (layers) to explore missing links in the sparse layer. By selectively fusing relevant information from other layers, we learn relevant representations that capture the characteristics of the sparse layer. Additionally, we incorporate node similarity information based on the relevant representation to enhance the graph structure of the sparse layer. By augmenting the graph structure, our approach improves the representation learning process and enables a more comprehensive exploration of relational patterns and connections within the sparse layer. Experimental evaluations on three real-world datasets demonstrate the effectiveness of our proposed SmGNN approach.
Submission Length: Regular submission (no more than 12 pages of main content)
Assigned Action Editor: ~Yujia_Li1
Submission Number: 2128
Loading