Figure 4.

Method comparison for real-world networks with respect to (a) change point classification, (b) Q_P, (c) Q_P (for different numbers of segments l for Senate network), and (d) running time (logarithmic scale). In panel (a), only networks with known ground truth change points are shown, since otherwise change point classification cannot be computed. In panel (c), Q_P scores of 1) SCOUT’s solutions for different numbers of segments l and 2) the solutions of the existing methods are shown. Specifically, for SCOUT, the line shows its Q_P score when solving the Q_P-based constrained segment community detection (CSCD) problem while varying the number of segments. For each of the existing methods, the mark shows Q_P score of the given method’s solution, with its position along the x-axis corresponding to the number of segments l in the solution. Equivalent results to those in panel (c) for the remaining networks are shown in Supplementary S21. In panel (d), the striped bar indicates that GHRG could not be run on AMD Hope network due to its high computational complexity.