Abstract
Centrality measures are crucial for network analysis, offering insights into node importance within complex networks. However, their accuracy is often affected by observational errors and incomplete data. This study investigates how sampling biases systematically impact centrality measures. We simulate six types of biased down-sampling, transitioning networks from dense to sparse states, using the initial network as the 'ground truth.' Changes in centrality values reveal the robustness of these measures under various sampling scenarios across synthetic and biological networks. Our results show that in synthetic networks, some sampling methods consistently exhibit higher robustness, particularly in scale-free networks. For biological networks, protein interaction networks are the most robust, followed by metabolite, gene regulatory, and reaction networks. Local centrality measures generally show greater robustness, while global measures are more heterogeneous and less reliable. This study highlights the limitations of centrality measures under sampling biases and informs the development of more robust methodologies.