Fig. 6: The robustness of resilience inference against observational noises.
From: Deep learning resilience inference for complex networked systems
We employ the same datasets, data split ratios, the number of initial conditions considered, and dynamics parameter settings as in Fig. 4. In contrast, we introduce observational noises into their node activities or network topologies. We repeat all experiments with 35 different random seeds, corresponding to different inference model initialization and split for datasets. a–c Noisy node activity: we evaluate the models' performance when various intensity levels of Gaussian noises are added to the observed node activity trajectories. Higher noise-to-signal ratio indicates stronger noise. GBB and SDR are left out in this comparison because they are not affected by node activity trajectories. Instead, GBB and SDR rely on explicit network dynamic equations that are not available to other methods. d–f Missing link: we evaluate the models' performance on networked systems with 5 ~ 30% links randomly removed. g–i Spurious link: we randomly add 5 ~ 30% links to pairs of nodes that are not originally connected and evaluate the models' performance on the perturbed networked systems. Centers represent the average values, and shadings represent the standard deviation (n = 35 with different random seeds).
