Representation learning on heterogeneous information networks

Abstract

With the superiority of representation learning with deep learning being well demonstrated across various fields, representation learning on graphs has gained heated attention, leading to a wide range of Intriguing graph embedding models and techniques being developed and published. Moreover, with recent advancements in generative adversarial learning, the fundamental idea of combining generative adversarial learning and graph representation learning has arisen and proven useful. This final year project focuses on critical review and analytical and empirical study of an existing approach HeGan [11] which combines representation learning on heterogeneous information network with generative adversarial learning. Through reviewing and analytical study on the existing researches, shortcomings of the HeGan framework are identified and some modifications have been proposed to address them. Furthermore, through extending HeGan framework and conducting experiments on benchmark datasets, the empirical study shows some advances beyond past research by demonstrating the proposed extended framework outperforms the existing framework under certain condition.