.png){kind=icon}
.png){kind=icon}
Return
CTRL: Continuous-time representation learning on temporal heterogeneous information network
.png){kind=icon}
PRE
AI
DOI：10.1016/j.knosys.2026.115514.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Temporal representation learning on heterogeneous graphs is critical for scalable deep learning on evolving heterogeneous information networks (HINs). However, most existing approaches lack inductive capabilities, making them ineffective at handling new nodes or edges. Furthermore, prior temporal graph embedding methods are typically trained via temporal link prediction to simulate link formation, overlooking the evolution of high-order topological structures. To address these limitations, we propose CTRL, a novel Continuous-Time Representation Learning model specifically designed for temporal HINs. This model integrates three key components into a single GNN layer to preserve both heterogeneous node features and temporal structures: (i) a heterogeneous attention mechanism for measuring semantic correlations between nodes; (ii) an edge-based Hawkes process for capturing temporal influence across heterogeneous nodes; and (iii) dynamic centrality for reflecting the dynamic importance of each node. In addition, we introduce a new training strategy that uses future event prediction (in the form of subgraphs) to capture the evolution of high-order network structures. Extensive experiments on three benchmark datasets show that our CTRL model significantly outperforms various state-of-the-art approaches. We also validate the effectiveness of our design through ablation studies and visualization analyses. Our codes and datasets will be released at Github: https://github.com/ycy89/CTRL.git .
Keywords:
Temporal representation learning
Heterogeneous information networks
Continuous-time modeling
Graph neural networks
High-order topological structures
Journal
K
IF:
7.6
Papers:
1.2W
Citations:
4.5W
View more organizations