Exploring convolutional auto-encoders for representation learning on networks

Abstract

A multitude of important real-world or synthetic systems possess network structure. Extending learning techniques such as neural networks to process such non-euclidean data is therefore an important direction for machine learning research. However, till very recently this domain has received comparatively low levels of attention. There is no straight forward application of machine learning to network data as machine learning tools are designed for $i.i.d$ data,  simple euclidean data or grids. To address this challenge the technical focus of this dissertation is on use of graph neural networks for Network Representation Learning (NRL) i.e. learning vector representations of nodes in networks. Learning vector embeddings of graph-structured data is similar to embedding complex data into low-dimensional geometries. After the embedding process is completed, drawbacks associated with graph structured data are overcome. The current inquiry proposes two deep learning auto-encoder based approaches for generating node embeddings. The drawbacks in existing auto-encoder approaches such as shallow architectures and excessive parameters are tackled in the proposed architectures using fully convolutional layers. Extensive experiments are performed on publicly available benchmark network data-sets to highlight the validity of this approach.