Link Prediction in Heterogeneous Collaboration Networks

Abstract

Traditional link prediction techniques primarily focus on the effect of potential linkages on the local network neighborhood or the paths between nodes. In this article, we study both supervised and unsupervised link prediction in networks where instances can simultaneously belong to multiple communities, engendering different types of collaborations. Links in these networks arise from heterogeneous causes, limiting the performance of predictors that treat all links homogeneously. To solve this problem, we introduce a new supervised link prediction framework, Link Prediction using Social Features (LPSF), which incorporates a reweighting scheme for the network based on nodes’ features extracted from patterns of prominent interactions across the network. Experiments on coauthorship networks demonstrate that the choice for measuring link weights can be critical for the link prediction task. Our proposed reweighting method in LPSF better expresses the intrinsic relationship between nodes and improves prediction accuracy for supervised link prediction techniques. We also compare the unsupervised performance of the individual features used within LPSF with two new diffusion-based methods: Link Prediction using Diffusion Process (LPDP) and Link Prediction using Diffusion Maps (LPDM). Experiments demonstrate that LPDP is able to identify similar node pairs, even far away ones, that are connected by weak ties in the coauthorship network using the diffusion process; however, reweighting the network has little impact on prediction performance.