Distributed Processing and Management of Graph Neural Networks

点击翻译按钮获取中文摘要

There has been an ever-increasing need for analysing, training and inferring over massive volumes of graph-structured data. Graphs like social interaction networks, communication networks, and biological networks tend to get quite massive in size and require lots of memory and computational power to operate. Such requirements are hardly met by a single dedicated machine. In addition to that, more and more systems, nowadays, are becoming latency-critical. As real-time recommendation systems prevail, we would like to infer over incoming data in a matter of milliseconds which becomes a challenge as the amount of data grows. Over the past years, Graph Neural Networks have been the predominant framework for making inferences over graphs. However, existing research mostly focuses on static graphs which can be fit into a single machine. Moreover, these models are trained over entire graph datasets, which become intractable for latency-critical systems. There are several challenges in building scalable Graph Neural Networks with sufficiently low latency and high throughput. There is a technological gap between the academic research and large-scale utilization of proposed Graph Neural Networks due to the given constraints. Hence, this project aims to develop a scalable and unified framework for Graph Neural Networks to handle large-scale evolving data with low latency. In particular, we plan to 1. Explore approaches for partitioning evolving graph data to a cluster of machines. 2. Develop new methods for incremental learning over the incoming data. 3. Provide a unified framework for distributed training and inferring with Graph Neural Networks. Alignment with EPSRC research themes: Artificial intelligence and robotics, Engineering, Artificial intelligence and robotics, Information and communication technologies (ICT).