Skip to main content

Table 3 Rand Index, Purity and Average F1 scores against ground truth; the number of clusters is set to the number of ground truth classes

From: Hypergraph clustering by iteratively reweighted modularity maximization

  Citeseer Cora MovieLens TwitterFootball Arnetminer
(a) Rand Index scores; number of clusters set to the number of ground truth classes
hMETIS 0.6891 0.7853 0.5028 0.7697 0.3116
PaToH 0.7312 0.7208 0.4984 0.7618 0.1820
Spectral 0.7369 0.3117 0.4812 0.7765 0.3762
Zhou-Spectral 0.8267 0.5845 0.5006 0.9112 0.3851
Louvain - 0.7096 0.4982 - 0.4198
NDP-Louvain 0.8197 0.8441 0.5119 - 0.5359
IRMM 0.8245 0.889 0.5347 - 0.5506
(b) Cluster purity scores; number of clusters set to the number of ground truth classes
hMETIS 0.5249 0.6359 0.6914 0.2354 0.2984
PaToH 0.5724 0.6498 0.7139 0.2419 0.2391
Spectral 0.4839 0.5819 0.7294 0.7815 0.5169
Zhou-Spectral 0.5374 0.6115 0.742 0.8191 0.5827
Louvain - 0.7136 0.7364 - 0.4837
NDP-Louvain 0.7495 0.7441 0.7429 - 0.5968
IRMM 0.7732 0.779 0.7737 - 0.6173
(c) Average F1 scores; number of clusters set to the number of ground truth classes
hMETIS 0.1451 0.2611 0.4445 0.3702 0.3267
PaToH 0.071 0.1799 0.3239 0.1036 0.2756
Spectral 0.2917 0.2305 0.2824 0.4345 0.387
Zhou-Spectral 0.3614 0.2672 0.3057 0.5377 0.4263
Louvain - 0.2725 0.2874 - 0.4587
NDP-Louvain 0.3491 0.3314 0.3411 - 0.4948
IRMM 0.441 0.3966 0.4445 - 0.5299
  1. Citeseer, Cora, Movielens, TwitterFootball, and Arnetminer have 6, 7, 2, 20, and 10 classes, respectively. On some datasets, the Louvain and IRMM method return fewer clusters than the number of ground truth classes. In such cases, we do not report the results and leave the entries as “-."
  2. Best performance in each column is boldfaced