Scholarlink

<p>Graph-based analysis plays a crucial role in understanding complex networks by uncovering hidden patterns, structural properties, and dynamic evolution. In this research, an Influential Node Ranking method using deep Autoencoder (INRA) is proposed to compute influential score of nodes. This method integrates feature extraction, dimensionality reduction, deep autoencoder for latent representations and adjacency loss to enhance ranking accuracy of nodes in the complex systems. Initially, the structural features are extracted from the adjacency matrix and apply principal component analysis (PCA) for dimensionality reduction. The PCA transforms potential features into smaller set by identifying underlying pattern, and it preserves the original information. The obtained reduced feature matrix is then multiplied by the normalized adjacency matrix and resultant matrix given as input to an autoencoder. The autoencoder generates matrix of latent representations of significant node features. Further, L2-norm is applied on the matrix which is generated by autoencoder to obtain influential score of the nodes and the nodes are ranked according to their decreasing influential score. The proposed model’s efficiency is evaluated by computing reconstruction and adjacency loss, ensuring meaningful feature preservation. The performance of proposed INRA model is evaluated using accuracy, precision, recall and <i>F</i>1-score matrices in four real-world datasets. The INRA method is scalable and demonstrates improved performance on large datasets by enhancing accuracy with extensive training data. The comparative analysis reveals that INRA outperforms existing methods, achieving 93.32% accuracy, 82.66% precision and 71.11% recall on the CA-HepTh dataset. Similarly, in the CA-GrQc dataset, it attains 86.45% accuracy and 80.53% precision, while in the Human Protein (Vidal) dataset, it achieves 85.63% accuracy and 92.26% precision. Also, in the Hamsterster Friendships dataset, INRA demonstrates accuracy 93.6%, reinforcing its robustness, particularly in citation networks. These results confirm that our method effectively captures network structures and ranks influential nodes with high reliability. The validation of proposed model is performed by simulation-based SIR model and Kendall’s tau correlation measures.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

INRA: deep autoencoder-based node influence ranking method in complex networks

  • 期刊论文

  • 发表日期：

    2025年6月1日

查看全文
  • Asmita Patel,
  • Buddha Singh

摘要

Graph-based analysis plays a crucial role in understanding complex networks by uncovering hidden patterns, structural properties, and dynamic evolution. In this research, an Influential Node Ranking method using deep Autoencoder (INRA) is proposed to compute influential score of nodes. This method integrates feature extraction, dimensionality reduction, deep autoencoder for latent representations and adjacency loss to enhance ranking accuracy of nodes in the complex systems. Initially, the structural features are extracted from the adjacency matrix and apply principal component analysis (PCA) for dimensionality reduction. The PCA transforms potential features into smaller set by identifying underlying pattern, and it preserves the original information. The obtained reduced feature matrix is then multiplied by the normalized adjacency matrix and resultant matrix given as input to an autoencoder. The autoencoder generates matrix of latent representations of significant node features. Further, L2-norm is applied on the matrix which is generated by autoencoder to obtain influential score of the nodes and the nodes are ranked according to their decreasing influential score. The proposed model’s efficiency is evaluated by computing reconstruction and adjacency loss, ensuring meaningful feature preservation. The performance of proposed INRA model is evaluated using accuracy, precision, recall and F1-score matrices in four real-world datasets. The INRA method is scalable and demonstrates improved performance on large datasets by enhancing accuracy with extensive training data. The comparative analysis reveals that INRA outperforms existing methods, achieving 93.32% accuracy, 82.66% precision and 71.11% recall on the CA-HepTh dataset. Similarly, in the CA-GrQc dataset, it attains 86.45% accuracy and 80.53% precision, while in the Human Protein (Vidal) dataset, it achieves 85.63% accuracy and 92.26% precision. Also, in the Hamsterster Friendships dataset, INRA demonstrates accuracy 93.6%, reinforcing its robustness, particularly in citation networks. These results confirm that our method effectively captures network structures and ranks influential nodes with high reliability. The validation of proposed model is performed by simulation-based SIR model and Kendall’s tau correlation measures.