Using Link Prediction to Estimate the Collaborative Influence of Researchers

Abstract

The influence of a particular individual in a scientific collaboration network could be measured in several ways. Estimating influence commonly requires calculating computationally costly global measures, which may be impractical on networks with hundreds of thousands of vertices. In this paper, we introduce new local measures to estimate the collaborative influence of individual researchers in a collaboration network. Our approach is based on the link prediction technique, and its underlying rationale is to assess how the presence/absence of a researcher affects the link prediction outcome in the network as a whole. It is natural to assume that the absence of a researcher with strong influence in the network will cause negative impact in the correct link prediction. Scientists are represented as vertices in the collaboration graph, and a vertex removal and corresponding link prediction process are performed iteratively for all vertices, each vertex being handled independently. The SVM supervised learning model has been adopted as link predictor. The proposed approach has been tested on real collaboration networks relative to multiple time periods, processing the networks in order to assign more relevance to recent than to older collaborations. The experimental tests suggest that our measure of impact on link prediction has high negative correlation with standard vertex importance measures such as between ness and closeness centrality.