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

Area of Science:

• Complex Networks
• Machine Learning
• Graph Theory

Background:

• Temporal networks are crucial in various fields but are often modeled with restrictive dynamic graph neural networks (DGNNs).
• Existing DGNNs struggle with multiple evolutionary mechanisms, limiting their ability to capture diverse spatiotemporal dynamics and leading to suboptimal performance and interpretability.
• The single-evolutionary-path architecture in current methods fails to explore rich evolutionary paths inherent in complex temporal networks.

Purpose of the Study:

• To propose a novel theoretical model, the multiway autoregressive (MARS) model, for characterizing 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 dynamics and improve the performance and interpretability of temporal link prediction.

Main Methods:

• Developed the multiway autoregressive (MARS) model to capture dependencies within and across factors driving network evolution.
• Designed a general DGNN framework (MAN) with a 2-D architecture representing spatiotemporal evolutionary dependencies.
• Incorporated spatialwise, temporalwise, and cross-spatiotemporal evolutionary operators into the MAN framework.

Main Results:

• Experimental validation using a novel DGNN incorporating GCN, GRU, and TD-GCN demonstrated superior performance.
• The proposed approach achieved state-of-the-art results in temporal link prediction on both synthetic and real-world datasets.
• The MAN framework effectively captures diverse evolutionary paths and enhances spatiotemporal interaction modeling.

Conclusions:

• The multiway autoregressive (MARS) model and multiway autoregressive network (MAN) framework offer a significant advancement in modeling complex temporal networks.
• This approach overcomes the limitations of single-evolutionary-path DGNNs, improving performance and interpretability.
• The proposed method provides a flexible and effective solution for temporal link prediction across various domains.