Link Prediction Measures in Various Types of Information Networks: A Review

Abstract

An information network is represented as a graph where nodes represent entities and edges represent interactions between nodes. There can be multiple types of nodes and edges in such networks giving rise to homogeneous, multi-relational and heterogeneous networks. Link prediction problem is defined as predicting edges that are more likely to be formed in the network at a future time. Many measures have been proposed in the literature for homogeneous networks. Extensions of many of these measures to heterogeneous networks are not available. Further, the measures need to be redefined in order to utilize the weight and time information available with the interactions. In this work, along with the logical grouping of the measures as topological, probabilistic and linear algebraic measures for all types of networks, we fill the gaps by defining the measures where ever they are not available in the literature. The empirical evaluation of each of these measures in different types of networks on the DBLP benchmark dataset is presented. An overall improvement of 12% is observed in prediction accuracy when temporal and heterogeneous information is efficiently utilized.