Improving network community structure with link prediction ranking

Abstract

Community detection is an important step of network analysis that relies on the correctness of edges. However, incompleteness and inaccuracy of network data collection methods often cause the communities based on the collected datasets to be different from the ground truth. In this paper, we aim to recover or improve the network community structure using scores provided by different link prediction techniques to replace a fraction of low ranking existing links with top ranked predicted links. Experimental results show that applying our approach to different networks can significantly refine community structure. We also show that predictions of edge additions and persistence are confirmed by the future states of evolving social networks. Another important finding is that not every metric performs equally well on all networks. We observe that performance of link prediction ranking is correlated with certain network properties, such as the network size or average node degree.