Predictive representation learning in motif-based graph networks

Link prediction is an important task for analyzing social networks which also has other applications such as bioinformatics and e-commerce. Network representation learning (NRL), which can significantly enhance the performance for link prediction, has attracted much attention in recent years. However, the existing NRL methods mainly focus on observed network structures without considering hidden prediction knowledge in the representation space. Meanwhile, some random walk based NRL methods are dissatisfactory to learn link knowledge in dense networks with large scales. In this paper, we propose a predictive representation learning (PRL) model, which unifies node representations and motif-based structures, to improve prediction ability of NRL. We firstly enhance node representations based on motif-biased random walks and then employ L2-SVM to learn motif-connected node-pairs. By jointly optimizing two objectives of existent and nonexistent edges representations, we preserve more information of nodes in representation space based on supervised learning. To evaluate the performance of our proposed model, we implement experiments on 5 real data sets. Simulation results illustrate that our proposed model achieves better link prediction performance compared with other state-of-the-arts methods.