Adaptive Inference Neuro-Fuzzy Driven Embedding Fusion for Improving Link Prediction Through Graph Neural Network
Link prediction has long been regarded as a fundamental problem in networked data analysis and mining, owing to its importance in a wide range of real-world applications, including social network analysis, recommendation systems, and biological network modeling. Over the years, numerous approaches have been proposed to address this problem, ranging from similarity-based heuristics and probabilistic models to dimensionality reduction and deep learning (DL)-based representation learning techniques. These methods aim to capture latent relational patterns among nodes in order to infer missing or potential links. Recently, graph neural networks (GNNs) have emerged as a powerful DL paradigm for graph-structured data representation learning and have demonstrated strong performance in complex link prediction tasks. In particular, multi-view GNN-based approaches have gained increasing attention by integrating different GNN architectures to capture complementary structural perspectives, such as local neighborhood proximity and global graph topology. This strategy enables the learning of richer node embeddings and improves task-driven prediction performance. However, the integration of multiple GNN layers and architectures often introduces feature uncertainty and noise amplification, which can degrade model stability and predictive accuracy. To address these challenges, this paper proposes AFGNN, a novel framework that integrates adaptive neuro-fuzzy inference with multi-view graph neural learning. By leveraging a robust fuzzy neural architecture with an adaptive inference engine, our proposed AFGNN model dynamically regulates the fusion of view-specific embeddings, enhances reasoning capability, and suppresses uncertainty during representation learning. Extensive experiments on benchmark graph datasets demonstrate the effectiveness and superiority of AFGNN compared with state-of-the-art GNN-based link prediction models.