Measure | Removed anomalous edges | \(\mu _0\) |
|---|
0.01 | 0.05 | 0.5 |
|---|
CS | Yes | \(0.837 \pm 0.015\) | \(0.834 \pm 0.020\) | \(0.838 \pm 0.014\) |
No | \(0.767 \pm 0.039\) | \(0.766\pm 0.040\) | \(0.768 \pm 0.040\) |
AUC | Yes | \(0.882 \pm 0.018\) | \(0.883 \pm 0.019\) | \(0.888 \pm 0.020\) |
No | \(0.953 \pm 0.058\) | \(0.936 \pm 0.056\) | \(0.922 \pm 0.045\) |
- We present the ability of ACD in community detection, represented by cosine similarity (CS). Moreover, we validate the model by measuring the AUC in link prediction tasks. The results are robust with respect to the initial values of \(\mu\). By removing the edges detected as anomalies, the community detection task is improved. Here, \(\pi = 10^{-3}\). The CS errors are calculated by averaging over 20 runs of the edge removal routine. To estimate AUC we perform 5-fold cross validation and report the averages and standard deviations over the fivefolds
