Linear clustering process on networks: a comparative study

Abstract

Abstract By a thorough performance comparison, we compare the recently proposed, operator-based linear clustering process on a network with classical, existing clustering algorithms. The linear clustering process produces clusters or partitions based on the eigenstructure of a linear operator on a graph that replaces nodes to ‘more natural’ positions by attractive and repulsive forces. Synthetic benchmarks, along with real-world networks possessing or lacking a known community structure, are considered. Our comparative analysis demonstrates that our linear clustering process generates superior partitions compared to the algorithms assessed in most instances, while of comparable computational complexity with the simplest existing clustering algorithms.