Figure 6:

The edge prior accurately assigns the probability of edge existence. (A) Calibration curves for full network reconstruction of 20 networks from Hetionet. For every unique predictor value on the horizontal axis, the fraction of node pairs with that predictor value having an edge in the network is shown on the vertical axis. The permutation-based edge prior’s calibration was superior to the other 2 strategies based on degree. (B) Calibration curves for sampled network reconstruction. The edge prior shows superior calibration in the 20 Hetionet networks. (C) Individual Hetionet edge-type calibration estimated by the 2-component decomposition of the Brier score, in which lower scores indicate better calibration. The edge prior has excellent calibration in unsampled and sampled networks, and each considered method is sensitive to shifts in the degree distribution.