Abstract:
Detecting an accurate community structure is a central problem in network analysis. With the popularity of social networking sites, a community detection algorithm's performance and energy efficiency are equally vital for minimizing the data centre's high operational cost for running this algorithm. There are several metrics for estimating the accuracy of community detection. Still, prior work has shown that a vertex-centric metric, permanence, provides a high precision estimate of a community structure compared to existing approaches [6]. However, permanence calculation is computationally heavy, and there has been no prior study on parallelizing this community detection algorithm. Genperm is an metric which is generalization of permanence to detect overlapping communities in an network. In this paper we have made three major contribution i) Amoeba: a high performance energy e cient multi-core community detection algorithm ii) Parallel multi-core energy e cient im- plementation of MaxGenperm to detect overlapping communities in an network iii) Detecting disjoint communities using GPUs by Parallelizing Maxperm on GPUs, We have provided exper- imental evaluation for all three contribution along with our insights and motivation.