Skip to main content

Table 6 Gap between the best NDCG value and the NDCG of the algorithm corresponding to the best measure value

From: Unsupervised evaluation of multiple node ranks by reconstructing local structures

   Structural Measures Local Structure Measures  
Network Examples Cond/nce Density Modularity LinkCC LinkAUC HopAUC AUC
Amazon 0.1% 24% 23% 13% 0% 0% 0% 0%
Amazon 1% 5% 4% 5% 4% 1% 1% 1%
Amazon 10% 9% 4% 7% 2% 2% 6% 0%
Amazon 25% 10% 14% 19% 2% 6% 6% 1%
Amazon 50% 35% 13% 17% 12% 4% 4% 2%
DBLP 0.1% 8% 5% 3% 7% 3% 4% 0%
DBLP 1% 4% 5% 3% 7% 15% 0% 0%
DBLP 10% 3% 0% 9% 5% 0% 0% 4%
DBLP 25% 7% 3% 9% 6% 10% 15% 9%
DBLP 50% 9% 20% 11% 5% 7% 7% 21%
BlockModel 0.1% 32% 18% 32% 8% 8% 8% 0%
BlockModel 1% 32% 14% 0% 39% 9% 10% 0%
BlockModel 10% 48% 20% 11% 51% 31% 16% 0%
BlockModel 25% 45% 17% 24% 27% 24% 24% 0%
BlockModel 50% 27% 38% 0% 28% 14% 2% 12%
CiteSeer 0.1% 10% 3% 3% 1% 2% 2% 3%
CiteSeer 1% 3% 4% 3% 4% 2% 4% 3%
CiteSeer 10% 10% 5% 10% 2% 6% 6% 6%
CiteSeer 25% 12% 5% 12% 3% 8% 8% 6%
CiteSeer 50% 7% 10% 7% 2% 3% 3% 1%
PubMed 0.1% 6% 13% 2% 12% 19% 18% 0%
PubMed 1% 6% 9% 7% 6% 12% 7% 1%
PubMed 10% 8% 6% 4% 19% 9% 9% 0%
PubMed 25% 12% 8% 4% 29% 4% 4% 0%
PubMed 50% 7% 11% 5% 22% 4% 3% 0%
Number of gaps ≤9% 13 13 17 16 19 20 23
Number of gaps ≤13% 18 17 22 18 21 20 24
  1. Gaps closer to zero mean that the NDCG found when selecting the algorithm optimizing the respective measure is close to the max NDCG between all algorithms. The smallest gap among unsupervised measures in each experiment setting is bolded