Multiway Autoregressive Network: A Dynamic Graph Representation Framework for Temporal Link Prediction

Area of Science:

• Complex networks
• Graph theory
• Machine learning

Background:

• Temporal networks in social and traffic systems exhibit complex evolutionary mechanisms.
• Existing dynamic graph neural networks (DGNNs) use a single-path architecture, limiting exploration of diverse evolutionary dynamics and hindering performance.
• Current DGNNs often passively model spatiotemporal interactions, impacting interpretability and predictive accuracy.

Purpose of the Study:

• To propose a novel theoretical model, the multiway autoregressive (MARS) model, to characterize multiple evolutionary paths in temporal networks.
• To develop a general dynamic graph neural network (DGNN) framework, the multiway autoregressive network (MAN), based on the MARS model.
• To enhance the modeling of spatiotemporal dependencies and improve temporal link prediction.

Main Methods:

• Developed the multiway autoregressive (MARS) model to capture dependencies within and across evolutionary factors.
• Designed the multiway autoregressive network (MAN) framework, visualizing network architecture as a 2-D diagram for spatiotemporal evolutionary dependencies.
• Incorporated spatialwise, temporalwise, and cross-spatiotemporal evolutionary operators within the MAN framework.

Main Results:

• The proposed MAN framework enables the configuration of diverse DGNNs by adjusting evolutionary operators.
• Experimental validation using a DGNN with GCN, GRU, and TD-GCN components demonstrated significant improvements.
• Achieved state-of-the-art performance in temporal link prediction on both synthetic and real-world temporal networks.

Conclusions:

• The MARS model and MAN framework effectively address limitations of single-evolutionary-path DGNNs.
• The proposed approach offers enhanced flexibility and interpretability in modeling complex temporal network dynamics.
• The framework facilitates the development of advanced DGNNs for superior temporal link prediction across various domains.