PROJECT TITLE :

Domain-adversarial Network Alignment

ABSTRACT:

The task of aligning networks is an important one in many different types of industries. A great number of previously published works rely on representation learning to complete this task; however, they do so without first removing the domain representation bias that is induced by domain-dependent features, which results in subpar alignment performance. In this paper, a unified deep architecture known as DANA is proposed as a means of obtaining a domain-invariant representation for the purpose of network alignment through the use of an adversarial domain classifier. To be more specific, we make use of Graph Convolutional Networks to carry out network embedding in accordance with the domain adversarial principle. This is done with only a limited number of observed anchors. After that, the framework for Semi-Supervised Learning is optimized by simultaneously maximizing the loss of a domain classifier and a posterior probability distribution of observed anchors. This is done in order to maximize the likelihood of observed anchors. Additionally, we develop a few variants of our model, such as weight-sharing for directed networks, direction-aware network alignment, and simplification of parameter space. Experiments conducted on three different datasets taken from real-world Social Networks have shown that the proposed methods achieve state-of-the-art levels of alignment accuracy.