Fig. 8: The 5 highest Gini importance scores for link prediction in each of the 19 real-world networks.
From: Sequential stacking link prediction algorithms for temporal networks
a shows the most important features for the completely-unobserved setting and b shows the partially observed setting. Acronyms for network features (detailed in SI Section A) include: PageRank (PR), the dot product of columns i and j in truncated low-rank approximation (dLRA-approx), average neighbor degree (AND), shortest path (SP), betweenness centrality (BC), degree centrality (DC), truncated low-rank approximation (LRA-approx), common neighbors (CN), low-rank approximation (LRA), Katz centrality (KC), closeness centrality (CC), the Adamic-Adar index (AA), preferential attachment (PA), eigenvector centrality (EC), resource allocation score (RA), network transitivity (NT), local clustering coefficients (LCC), Jaccard’s coefficient (JC), local centralities (LC), network diameter (ND), Leichet-Holme-Newman index (LHN), average of entries of i and j’s neighbors in low rank approximation (mLRA). The vertical lines are provided to highlight the visually apparent separation in Fig. 7 of the 8 datasets on the right here from the others.
