Fig. 2

TL-BSLA performs consistently better than SCA in four example systems. The true positive rate (TPR), false positive rate (FPR), and positive predictive value (PPV) are shown for four representative networks. Black lines show performance of TL-BSLA, blue lines show performance of SCA. Dashed lines represent calculations without considering the direction of connections when assessing their correctness. TL-BSLA is almost universally better than SCA, with the exception of TPR for the Asia and Alarm networks where the directionality is not accounted for in assessing correctness. In these cases, the much higher FPR of SCA outweighs its potentially better coverage of true positives, as evidenced in the superior PPV curves for TL-BSLA. For PPV, all performance metrics across all networks (directed and undirected) are statistically significant (p < 0.05, two-tailed t-test) except for the 50 and 150 sample sizes for the Asia network for the undirected case. Error bars are one standard deviation