School of
Information Technology and Electrical Engineering

Speaker: Mr Mubashir Imran
Seminar Date: Thu, 13/02/2020 - 13:00
Venue: 78-631; Data Science MM Lab
Host: Dr Hongzhi Yin

Seminar Type:  PhD Confirmation Seminar

Abstract: 

Network embedding aim to learn vector representation of vertices, that preserve both network structure and properties. Most state-of-the-art methods can effectively process network data that comprise thousands of nodes. However, with the advent of web 2.0 and web 3.0 most real-world datasets now are perpetually growing, having very large and complex network structure. Existing embedding methods fail to scale for such large networks, due to high computation demands. A few frameworks have been proposed, that extend existing methods to drive node embedding on large scale networks. These frameworks, update the global parameters iteratively or compress the network, while learning the vector representation. This generates node embedding at the cost of either communication overhead or embedding quality.

In this work, we explore a framework that embeds large-scale networks in a decentralized fashion, evading communication overheads. In our work, we propose a novel decentralized large-scale network embedding framework, called DeLNE. As the name suggests, DeLNE divides a network into smaller partitions and learn the vector representation in a distributed fashion, avoiding any communication overhead. Our proposed framework uses Variational Graph Convolution Auto-Encoders to embed structure and properties of each sub-network. Secondly, we propose an embedding aggregation mechanism, that captures the global properties of each node within a sub-network. Thirdly, we propose an alignment function, that reconciles the sub-networks embedding onto the same vector space. Due to the parallel nature of DeLNE, it scales well on large clustered environments. Extensive experimentation on realistic datasets shows that DeLNE produces high quality embedding and outperforms existing large-scale network embedding frameworks, in terms of both efficiency and effectiveness.

Biography: 

Mubashir Imran is a PhD candidate at DKE group under the supervision of Dr. Hongzhi Yin and A/Prof Helen Huang. He received his Master and Bachelor degree in Computer Science from Information Technology University and University of Central Punjab, Pakistan. His research interests include graph neural networks, information retrieval and big data.