LinkAUC: Unsupervised Evaluation of Multiple Network Node Ranks Using Link Prediction

An emerging problem in network analysis is ranking network nodes based on their relevance to metadata groups that share attributes of interest, for example in the context of recommender systems or node discovery services. For this task, it is important to evaluate ranking algorithms and parameters and select the ones most suited to each network. Unfortunately, large real-world networks often comprise sparsely labelled nodes that hinder supervised evaluation, whereas unsupervised measures of community quality, such as density and conductance, favor structural characteristics that may not be indicative of metadata group quality. In this work, we introduce LinkAUC, a new unsupervised approach that evaluates network node ranks of multiple metadata groups by measuring how well they predict network edges. We explain that this accounts for relation knowledge encapsulated in known members of metadata groups and show that it enriches density-based evaluation. Experiments on one synthetic and two real-world networks indicate that LinkAUC agrees with AUC and NDCG for comparing ranking algorithms more than other unsupervised measures.