21. Accuracy on corrupted networks

21. Accuracy on corrupted networks

interactomes are incomplete, so the realistic test is ranking stability when the observed network is degraded. From the 1,200-gene subnetwork I removed 10%, 25% or 50% of edges uniformly at random (3 repeats each) and recomputed NetPert and betweenness on the degraded network only, comparing against the full-network ranking.

figure
Figure 22. Most-used shortest-path mediators in the Twist1 network.

NetPert's rho falls from 1.0 to 0.948, 0.889 and 0.775 at 10%, 25% and 50% removed; betweenness falls faster (0.886, 0.723, 0.520). NetPert's lead therefore grows with degradation — the reverse of the simulated-network finding. Top-20 overlap falls from 17.7 to 11.0 to 9.7.

Caveats. Only 3 repeats per level, so intervals are indicative; uniform removal understates study bias. This is still a stability result, not accuracy.