Evaluation of Correlation Measures for Computationally-Light vs. Computationally-Heavy Centrality Metrics on Real-World Graphs

We identify three different levels of correlation (pairwise relative ordering, network-wide ranking and prediction through linearity) that could be assessed between a computationally-light centrality metric and a computationally-heavy centrality metric for real-world networks. The Kendall's concordance-based correlation measure could be used to quantitatively assess how well we could consider the relative ordering of two vertices vi and vj with respect to a computationally-light centrality metric as the relative ordering of the same two vertices with respect to a computationally-heavy centrality metric. We hypothesize that the pair-wise relative ordering (concordance)-based assessment of the correlation between centrality metrics is the strictest of all three levels of correlation and claim that the Kendall's concordance-based correlation coefficient will be lower than the correlation coefficient observed with the more relaxed levels of correlation measures (prediction through linearity-based Pearson's product-moment correlation coefficient and the network wide ranking-based Spearman's correlation coefficient). We validate our hypothesis by evaluating the three correlation coefficients between two sets of centrality metrics: the computationally-light degree and local clustering coefficient complement-based degree centrality metrics and the computationally-heavy eigenvector centrality, betweenness centrality and closeness centrality metrics for a diverse collection of 50 real-world networks.