PPNE: Property Preserving Network Embedding

Network embedding aims at learning a distributed representation vector for each node in a network, which has been increasingly recognized as an important task in the network analysis area. Most existing embedding methods focus on encoding the topology information into the representation vectors. In reality, nodes in the network may contain rich properties, which could potentially contribute to learn better representations. In this paper, we study the novel problem of property preserving network embedding and propose a general model PPNE to effectively incorporate the rich types of node properties. We formulate the learning process of representation vectors as a joint optimization problem, where the topology-derived and property-derived objective functions are optimized jointly with shared parameters. By solving this joint optimization problem with an efficient stochastic gradient descent algorithm, we can obtain representation vectors incorporating both network topology and node property information. We extensively evaluate our framework through two data mining tasks on five datasets. Experimental results show the superior performance of PPNE.