Scholarlink

<p>Recently, various methods have emerged for predictive modeling across multiple domains to forecast the influence of nodes in complex networks. The existing machine learning based methods predict influential nodes in complex networks by using feature selection techniques for the nodes. However, these methods suffer from limitations in feature engineering. In contrast, Graph Convolutional Networks (GCNs) present a breakthrough in deep learning, offering an effective approach for analysing graph-structured data. This research work proposes the Influential node Prediction based on Graph Convolutional Network (Ip-GCN) method for complex networks, which integrates GCN and Long Short-Term Memory (LSTM) techniques. Initially, the proposed method preprocesses datasets by removing noisy nodes and constructing the adjacency matrix of the graph. The addition of the identity and adjacency matrices accounts for the self-contribution of individual nodes in the graph, and the degree matrix is derived from the updated adjacency matrix. Furthermore, the feature matrix is computed from the adjacency matrix to capture the connections of nodes and the hidden structure of the graph. The scaling of the adjacency matrix is performed using the inverse of the degree matrix, followed by normalization with the feature matrix. The resultant normalized matrix is provided as input to the GCN layer for embeddings. The LSTM layer is then applied to the embedded matrix to predict the influence of nodes. The comparative performance of Ip-GCN is evaluated against existing methods, including deep learning, machine learning and centrality-based methods. The Ip-GCN model significantly enhances performance across all five real-world datasets. It improves the F1 score by 9%, 29%, 14%, 9%, and 18% on Hamsterster Friendship, Human Protein (Vidal), CA-GrQc, CA-HepTh, and CA-CondMat respectively. In addition to that, it enhances accuracy by 5%, 1%, 0.4%, and 4% on Human Protein (Vidal), CA-GrQc, CA-HepTh, and CA-CondMat respectively. The predicted influential nodes by the Ip-GCN method are validated using the Susceptible Infected Recovered (SIR) simulation model .</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

Ip-GCN: a hybrid deep learning approach to predict influence in complex networks

  • 期刊论文

  • 发表日期：

    2025年3月28日

查看全文
  • Asmita Patel,
  • Naveen Kumar Singh,
  • Buddha Singh

摘要

Recently, various methods have emerged for predictive modeling across multiple domains to forecast the influence of nodes in complex networks. The existing machine learning based methods predict influential nodes in complex networks by using feature selection techniques for the nodes. However, these methods suffer from limitations in feature engineering. In contrast, Graph Convolutional Networks (GCNs) present a breakthrough in deep learning, offering an effective approach for analysing graph-structured data. This research work proposes the Influential node Prediction based on Graph Convolutional Network (Ip-GCN) method for complex networks, which integrates GCN and Long Short-Term Memory (LSTM) techniques. Initially, the proposed method preprocesses datasets by removing noisy nodes and constructing the adjacency matrix of the graph. The addition of the identity and adjacency matrices accounts for the self-contribution of individual nodes in the graph, and the degree matrix is derived from the updated adjacency matrix. Furthermore, the feature matrix is computed from the adjacency matrix to capture the connections of nodes and the hidden structure of the graph. The scaling of the adjacency matrix is performed using the inverse of the degree matrix, followed by normalization with the feature matrix. The resultant normalized matrix is provided as input to the GCN layer for embeddings. The LSTM layer is then applied to the embedded matrix to predict the influence of nodes. The comparative performance of Ip-GCN is evaluated against existing methods, including deep learning, machine learning and centrality-based methods. The Ip-GCN model significantly enhances performance across all five real-world datasets. It improves the F1 score by 9%, 29%, 14%, 9%, and 18% on Hamsterster Friendship, Human Protein (Vidal), CA-GrQc, CA-HepTh, and CA-CondMat respectively. In addition to that, it enhances accuracy by 5%, 1%, 0.4%, and 4% on Human Protein (Vidal), CA-GrQc, CA-HepTh, and CA-CondMat respectively. The predicted influential nodes by the Ip-GCN method are validated using the Susceptible Infected Recovered (SIR) simulation model .