CACE-Net: Cascade Coupling Effect for Link Prediction in Multi-layer Networks

Ruohan Yang, Muhammad Asif Ali, Anyu Xue, Junyang Chen, Ying Sha, Di Wang, Huan Wang

Keywords: Link prediction, multi-layer networks, cascade coupling effects, adversarial training.

Abstract: As social infrastructure networks become increasingly complex, the interdependence between different network layers has garnered significant attention. In real-world multi-layer networks, structural changes in one layer often trigger cascade coupling effects, where these changes propagate across layers and influence link formation in a chain-like manner. However, traditional link prediction methods typically treat each layer independently, overlooking these cross-layer dependencies. To address this, we propose CACE-Net: CAscade Coupling Effect for link prediction in multi-layer Networks. CACE-Net encomasses three key components: (i) Layer-wise representation extractor; (ii) Adversarial coupling representation encoder; and (iii) Adaptive fusion link predictor. Firstly, layer-wise representation extractor applies independent graph convolutions to model intra-layer structures. Next, the adversarial coupling representation encoder leverages adversarial training to learn latent cascading dependencies between replica nodes across layers. Finally, adaptive fusion predictor integrates intra-layer and cross-layer embeddings via an attention mechanism, effectively combining local and global information to enhance link prediction in the target layer. Experimental results on multiple real-world datasets show that CACE-Net outperforms state-of-the-art methods, achieving AUC improvements of up to 13.29%.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 12361