Fig. 3
From: Machine learning meets complex networks via coalescent embedding in the hyperbolic space
HD-correlation on popularity-similarity-optimization synthetic networks. a–i To validate the above-mentioned techniques, we generated 100 different synthetic networks for each combination of tuneable parameters of the PSO model (temperature T, size N, half of average degree m, power-law degree distribution exponent γ). Supplementary Fig. 24 offers an idea of the topological diversity of the synthetic networks generated fixing γ = 2.5 and tuning the other parameters, Supplementary Fig. 25 reports an analysis of the rich clubness of the networks, commented in Supplementary Discussion. In the results presented in the figures of this article, we used γ = 2.5, but we also ran the simulations for γ = 2.25 and 2.75, and the differences were negligible (results not shown). Here, the performance was evaluated as Pearson correlation between all the pairwise hyperbolic distances of the network nodes in the original PSO model and in the reconstructed hyperbolic space (HD-correlation). The plots report the average correlation and the standard error over the 100 synthetic networks that have been generated for each different parameter combination. The value one indicates a perfect correlation between the nodes’ hyperbolic distances in the original and reconstructed hyperbolic space. The plots show the results of different methods when both RA and EA are applied. The methods without EA are plotted in Supplementary Fig. 7. For each subplot, the value of HyperMap-CN for T = 0 is missing because the original code assumes T > 0
