Network Representation Learning with Its Applications

Abstract

Network Representation Learning (NRL) is a powerful technique for typical machine learning tasks in social and information networks, such as link prediction and node classification. NRL aims at learning an effective vector transformation so that nodes in a graph can be projected into a low-dimensional feature space. State-of-the-art NRL methods mainly focus on homogeneous graphs, i.e., nodes and links are of the same types. We have developed novel embedding learning algorithms in heterogeneous graphs that contain different types of nodes and links. The goal is to let the learned embedding preserve not only graph proximity between nodes but also relational and attributed information associated on nodes and links. Based on our proposed NRL on homogeneous and heterogeneous graphs, in this talk, we will present two recent studies. One is how to deliver a more effective recommender system through user identification behind shared accounts. The other is how to boost the performance of link prediction and node classification in both homogeneous and heterogeneous realms. I will also demonstrate the promising prediction power of our NRL models by the experiments conducted on six real-world network datasets.