Structured Sparse Linear Model for Social Trust Prediction

Social trust prediction aims at predicting the missing trust relations between online users. In this paper, we propose a novel and scalable structured sparse linear model for social trust prediction from a global neighborhood-based collaborative filtering perspective. We formulate the prediction problem as a set of independent linear regression problems regularized by pairwise elastic net, to automatically learn correlation coefficients between a user and its most similar neighbors. In order to deal with large-scale sparse social trust data, we utilize efficient hashing techniques and stochastic coordinate descent algorithm to cut down the computational cost of training model. The experimental results on three real-world data sets show that our approach can significantly outperform the other tested methods in terms of prediction quality and efficiency.