Improving sign prediction of network embedding by adding motif features

Abstract

Sign prediction plays an important role in friendship recommendation and trust relationship prediction, so it has attracted a lot of attention from network embedding researchers in the field of machine learning. However, these algorithms may not fully exploit local structural similarity for each pair of nodes in traditional network domain, and few researchers have explored whether network structure features can supplement these embedding methods to boost the performance of sign prediction. In this study, we integrate 3-node and 4-node motifs into four network embedding algorithms, that is, we fuse motif and embedding features for sign prediction in four real undirected signed networks, and the performance can be greatly improved (up to 79.1%). Furthermore, the performance of four popular embedding algorithms are close to each other after fusing motif features, which demonstrates that motif features can compensate for the information of negative edges ignored by unsigned embeddings. Adding motif features to network embedding helps to offset the shortcoming of current embedding algorithms, that is, the inability of uncovering local structure features in classical network domain.