A hybrid machine learning and centrality framework for key node identification in complex networks

Determining the most significant nodes in complex networks is more crucial for a broad spectrum of applications. In order to assess a node’s significance, traditional centrality metrics are largely based on the network’s structural configuration. However, the asymmetric link between a node’s structural features, including local information, and its functional value is not well captured by them. To overcome these issues, we propose a machine learning approach for node assessment that identifies the nonlinear interaction between network structure and node functionality. We implement a system that derives features for each node as a vector, by utilizing the traditional existing centrality measures like degree (D), betweenness (B), closeness (C), katz (K), clustering coefficient (CC), pagerank (PR), global relative change in average closeness (GRAC), global relative change in average clustering coefficient (GRACC), global relative change in average katz (GRAK), global relative change in average betweenness (GRAB), global relative change in average pagerank (GRAPR), and global relative change in average degree (GRAD). The system incorporates the infection rate as a key factor in contagion modelling, labelling each node by its verified spreading ability through Independent Cascade and SIR simulations. Our main objective is to comprehend the underlying relationship between a disease’s actual spreading capacity and its rate of infection using machine learning techniques. The machine learning model effectiveness is evaluated in two scenarios: (1) exhibits higher accuracy than traditional centrality measures when trained and tested on the same network, (2) while GRACC, GRAK, GRAPR, GRAB, GRAC, and GRAD outperform the machine learning techniques when trained data is from one network and tested data is from another network. The suggested machine learning method exhibits an accuracy of 25% higher than existing centrality approaches in two distinct scenarios, highlighting its potential in a wide range of applications.