Core Concepts
A parallel implementation of the Louvain algorithm, called GSP-Louvain, that effectively addresses the issue of internally-disconnected communities while achieving significantly faster processing speeds compared to existing alternatives.
Abstract
The content presents a parallel implementation of the Louvain algorithm, called GSP-Louvain, that addresses the problem of internally-disconnected communities. The key highlights are:
The Louvain algorithm, a widely used method for community detection, can sometimes produce internally-disconnected communities. To mitigate this issue, the authors propose the GSP-Louvain algorithm.
GSP-Louvain employs a Split Pass (SP) approach, where disconnected communities are identified and split after the local-moving phase in each iteration of the Louvain algorithm, using a parallel Breadth-First Search (BFS) technique.
Evaluated on a system with two 16-core Intel Xeon Gold 6226R processors, GSP-Louvain achieves a processing rate of 328M edges/s on a 3.8B edge graph, outperforming the original Leiden, igraph Leiden, and NetworKit Leiden algorithms by 341x, 83x, and 6.1x, respectively.
The communities identified by GSP-Louvain are of similar quality to the Leiden and igraph Leiden implementations, and 25% higher in quality than NetworKit Leiden.
GSP-Louvain exhibits a performance improvement rate of 1.5x for every doubling of threads, demonstrating good scalability.
Stats
On a 3.8B edge graph, GSP-Louvain achieves a processing rate of 328M edges/s.
GSP-Louvain outperforms the original Leiden, igraph Leiden, and NetworKit Leiden algorithms by 341x, 83x, and 6.1x, respectively.