MWTP: A heterogeneous multiplex representation learning framework for link prediction of weak ties

Area of Science:

• Complex Networks
• Network Science
• Data Mining

Background:

• Weak ties are vital for network connectivity and resilience but difficult to predict due to limited common neighbors.
• Multiplex networks, with multiple interaction layers, offer potential to overcome this challenge by leveraging cross-layer information.

Purpose of the Study:

• To propose a novel Graph Neural Network (GNN)-based framework for Multiplex Weak Tie Prediction (MWTP).
• To effectively learn and fuse information across different layers of multiplex networks for improved link prediction.

Main Methods:

• Developed a GNN-based representation learning framework (MWTP) utilizing both intra-layer and inter-layer aggregators.
• Intra-layer aggregation incorporates multi-order neighbor features.
• Inter-layer aggregation uses logit regression (MWTP-logit) or semantic voting (MWTP-semantic) for nodal-level attention, creating efficient and capable variants.

Main Results:

• MWTP frameworks significantly outperformed eleven baseline methods in predicting weak ties and all ties on real-world multiplex networks.
• Both MWTP variants demonstrated strong prediction performance, even with limited training data.
• MWTP-semantic showed stronger learning capabilities, while MWTP-logit offered greater implementation efficiency.

Conclusions:

• The proposed MWTP framework effectively leverages multiplex network structures for accurate weak tie prediction.
• This approach enhances network resilience and connectivity by improving the understanding of weak tie dynamics.
• MWTP offers a promising solution for link prediction in complex systems with multiple interaction types.