Signed Network Embedding Based on Noise Contrastive Estimation and Deep Learning

Abstract

Network embedding is a representation learning method to learn low-dimensional vectors for vertices of a given network, aiming to capture and preserve the network structure. Signed networks are a kind of networks with both positive and negative edges, which have been widely used in real life. Presently, the mainstream signed network embedding algorithms mainly focus on the difference between positive and negative edges, but ignore the role of empty edges. Considering the sparsity of signed networks, a signed network embedding algorithm NDSE is proposed based on noise contrastive estimation model and deep learning framework. The algorithm emphasizes the role of empty edges and optimizes a carefully designed objective function that preserves both local and global network structures. Empirical experiments prove the effectiveness of the NDSE on three real data sets and one signed network task.