Transition Propagation Graph Neural Networks for Temporal Networks

Area of Science:

• Graph Neural Networks
• Network Science
• Data Mining

Background:

• Temporal networks, such as social and transaction networks, exhibit dynamic node patterns.
• Existing graph mining methods like skip-gram and GNNs generate sequential node embeddings.
• Sequential modeling struggles with transition structures between node neighbors due to limited memory.

Purpose of the Study:

• To propose a novel method, Transition Propagation Graph Neural Networks (TIP-GNN), for encoding transition structures in temporal networks.
• To effectively model nodes' personalized patterns and capture node dynamics.
• To address the limitations of sequential modeling in handling complex temporal network dynamics.

Main Methods:

• TIP-GNN utilizes a bilevel graph structure, considering both explicit interaction graphs and implicit transition graphs derived from sequential interactions.
• Employs multistep transition propagation to encode transition structures.
• Incorporates bilevel graph convolution to distill neighborhood information.

Main Results:

• TIP-GNN achieved significant improvements in temporal link prediction accuracy, with up to 7.2% enhancement on various temporal networks.
• Extensive ablation studies confirmed the effectiveness of the transition propagation module.
• The method demonstrates superior performance in capturing complex node dynamics and transition structures.

Conclusions:

• TIP-GNN effectively encodes transition structures in temporal networks, outperforming existing sequential methods.
• The proposed approach offers a robust framework for analyzing dynamic patterns in temporal network data.
• The findings highlight the importance of considering transition structures for accurate temporal link prediction.