SocNL: Bayesian Label Propagation with Confidence

Abstract

How can we predict Smith’s main hobby if we know the main hobby of Smith’s friends? Can we measure the confidence in our prediction if we are given the main hobby of only a few of Smith’s friends? In this paper, we focus on how to estimate the confidence on the node classification problem. Providing a confidence level for the classification problem is important because most nodes in real world networks tend to have few neighbors, and thus, a small amount of evidence. Our contributions are three-fold: (a) novel algorithm; we propose a semi-supervised learning algorithm that converges fast, and provides the confidence estimate (b) theoretical analysis; we show the solid theoretical foundation of our algorithm and the connections to label propagation and Bayesian inference (c) empirical analysis; we perform extensive experiments on three different real networks. Specifically, the experimental results demonstrate that our algorithm outperforms other algorithms on graphs with less smoothness and low label density.