Link Weight Prediction for Directed WSN Using Features from Network and Its Dual

Abstract

Link prediction problem in social networks is a very popular problem that has been addressed as an unsupervised as well as supervised classification problem. Recently a related problem called link weight prediction problem has been proposed. Link weight prediction on Weighted Signed Networks (WSNs) holds great significance as these are semantically meaningful networks. We consider two groups of features from the literature - edge-to-vertex dual graph features and fairness-goodness scores in order to propose a supervised framework for weight prediction that uses fewer features than those used in the literature. Experimentation has been done using three different feature sets and on three real world weighted signed networks. Rigorous assessment of performance using (i) Leave-one-out cross validation and (ii) N% edge removal methods has been carried out. We show that the performance of Gradient Boosted Decision Tree (GBDT) regression model is superior to the results presented in the literature. Further the model is able to achieve superior weight prediction scores with significantly lower number of features.