Cross-Network Structure Enhancement via Adaptive Coupling

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

Keywords: Cross-Network, Adaptive coupling, Structure Enhancement

Abstract:

Network structural enhancement seeks to improve the accuracy and reliability of real-world network representations by systematically detecting and inferring missing or potential links.Existing research primarily focuses on single networks, overlooking the interdependence of real-world systems. In practice, entities often span multiple networks—for example, users migrate and interact across social platforms, forming multiplex networks. Approaches considering multiplex networks typically use static weights or simple aggregation, failing to adaptively control the influence of each network at the sample level. This can introduce irrelevant information and cause negative transfer.
To address this, we introduce Adaptive Coupling for cross-Network structure Enhancement (ACNE), the first framework that leverages adaptive, sample-wise cross-network coupling for structure enhancement in multiplex networks. We first employ GNNs to obtain network-specific representations. Building upon this foundation, we introduce a generative–discriminative adversarial learning framework, and impose an adversarial weight perturbation in parameter space to approximate worst-case noise and stabilize the learned cross-network embeddings. To adaptively balance the contributions between target-specific and cross-network embeddings, we design a low-rank bilinear gated fusion module. In addition, a decorrelation regularizer is incorporated to minimize redundancy arising from overlapping communities. Extensive experiments on real-world multiplex networks show that our approach consistently surpasses existing baselines in link prediction, highlighting the effectiveness and practical value of adaptive cross-network structure enhancement.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 8891