Link Prediction Using Node Influence Based on Graph Neural Networks

Read full text Summarise with VIRAT
Authors: Madhusudhana Rao Baswani, Khyathisree Yarra, Prasanthi Boyapati, Eswaraiah Rayachoti, T. Jaya Lakshmi
Cited by: 0
References: 48

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Link prediction in complex networks aims to infer missing or future connections between nodes, a task crucial for understanding network evolution in domains such as social systems, biology, and recommender platforms. Traditional similarity-based approaches, which rely primarily on local neighborhood overlap, often fail to capture global structural properties and node importance, resulting in limited predictive accuracy in sparse or large-scale networks. To address these limitations, we propose a Link Prediciton based Centrality using Graph Neural Network (LP-GNN) framework that integrates centrality measures with similarity indices in a unified, parameterized learning paradigm. The proposed LP-GNN operates in three sequential stages. (i) Feature Construction: Edge features are enriched by a weighted fusion of structural similarity and node centrality measures, providing a comprehensive representation of both local and global graph information. (ii) Centrality-Aware Encoder: A modified message-passing mechanism is employed to learn node embeddings, where the influence of structurally important nodes is amplified, ensuring that critical topological information is effectively captured. (iii) Multi-Branch Decoder: Multiple complementary scoring mechanisms are integrated including dot-product similarity, a multilayer perceptron (MLP) to model non-linear interactions, and distance-based measures to generate robust estimates of link probabilities between node pairs. Our approach was evaluated across multiple real-world benchmark datasets, demonstrating its superiority over both traditional similarity-based methods and recent GNN-based models. Notably, it achieves improvements of up to 49% in Area Under the Receiver Operating Characteristic Curve (AUROC) and Area Under the Precision-Recall Curve (AUPR) over state-of-the-art baselines. The highlight the effectiveness of combining structural interpretability from centrality with the representational power of GNNs, offering a robust and scalable solution to the link prediction problem in complex networks.

Fields of Study & Subjects

Advanced Graph Neural Networks Complex Network Analysis Techniques Graph Theory and Algorithms

Publication details

Journal / source
IEEE Access
Volume
14
Pages
51732–51749
Publication date
1 January 2026
Language
EN