Link prediction on Twitter
Fig 2
Link prediction in 25%, 50% and 75% of the links in networks constructed from all the words in tweets of the emo-neta, emo-netb, emo-netc and emo-netd datasets.
Shown are the evaluation metric scores (see legend), namely the F1 score, the precision, and the area under the receiver operating characteristic curve (AUC), as obtained for seven different link prediction measures, namely common neighbors (CN), the Jaccard coefficient (JC), preferential attachment (PA), Adamic-Adar (AA), the resource allocation index (RA), selectivity (SE) and inverse selectivity (IS). The values of the F1 score and of precision are decreasing with the longitudinal growth of the networks (from 25% to 75%), while the AUC does better at retaining values regardless of the used percentage of links. The PA link prediction measure exposes the lowest link prediction potential on the emo-net dataset, this is regardless of the evaluation metrics used. See Table 2 and the main text for details.