Large Scale Graph Analytics for Communities Using Graph Neural Networks

Abstract

One of the challenging research areas in modern day computing is to understand, analyze and model massively connected complex graphs resulting due to highly connected networks because of newly accepted paradigm of Internet of Things. Patterns of interaction between nodes reflect a lot of information about nature of underlying network graph. The connectedness of nodes has been studied by several researchers to provide near optimal solution about topological structure of the graphs. This is more commonly known as community detection, which in mathematical and algorithmic terms is often referred to as graph partitioning. The study is broadly based on clustering of nodes, which share similar properties. Lower order connection patterns that detect communities at node and edge level are extensively studied. A wide range of algorithms has been studied to identify communities in large-scale networks. Spectral clustering, hierarchical clustering, Markov models, modularity maximization methods, etc have shown promising results in context to application domains under consideration. In this paper, the authors propose a neural network based method to identify the communities in large-scale networks. The study is broadly based on clustering of nodes, which share similar properties. This work is devoted to identify the efficacy of neural networks in community detection for large and complex networks in comparison to existing methods. The approach is motivated by neural network pipeline for data embedding into lower dimensional space, which is expected to simplify the task of clustering data into communities with inherent ability to learn between mapping and predicted communities.