Skip to main content
. 2019 Oct 30;5(10):eaax3770. doi: 10.1126/sciadv.aax3770

Fig. 2. Evaluation on a synthetic dataset with heterogeneous structures.

Fig. 2

(A and B) The performances of the proposed method are evaluated on the synthetic dataset 01_chang_pathbased provided in ClustEval (2). This includes groups with heterogeneous structures (i.e., a thin-elongated structure surrounding two globular-like structures). Our method, with a generic minimax path-cost function, partially improved the results produced by CDP and DBSCAN (A and B) with respect to the ground truth (GT). By contrast, by training a path classifier, the algorithm achieved full performances with an F1 score ≥0.99 and a Jaccard index ≥0.98. (C) Training paths composed of 25 examples of desired paths and 25 examples of undesired paths. (D) Path features corresponding to the density profile (vector including the density of the nodes in the path).