Abstract: Identifying influential nodes is vital for addressing the task of influence maximization in complex networks. Traditional methods typically assess a node’s influence using centrality measures that focus solely on network topology. However, a node’s influence is also affected by other factors, such as temporal activity and infection rates. Machine learning (ML) models offer a more nuanced approach by integrating multiple factors to assess a node’s influence potential. However, these ML models often classify nodes as either influential or not, without evaluating how well they propagate influence. To address these limitations, we have reformulated the problem of assessing node’s influence potential as a regression chore and developed the BO-CBoost model. Our model firstly generates a feature vector for every node reliant on both structural and temporal aspects of a node along with the infection rate. This feature vector is then used for model building and testing. We compared BO-CBoost against benchmark models like decision trees and linear regression, as well as state-of-art models such as LightGBM and XGBoost. Experiments conducted on three diverse real-world networks showed that BO-CBoost outperforms the baseline models. It achieved the most precise influence spread predictions, with a minimum average mean squared error of 0.056 and the lowest average absolute error of 0.118. Thus, BO-CBoost proved to be a more effective tool for predicting influence potential of nodes in complex networks.

Key words: influence potential, influential node identification, complex network, machine learning, social network analysis