Efficient Estimation of Graph Trussness

TL;DR

This paper introduces a faster approximation algorithm for estimating the graph's trussness, a measure of cohesive subgraphs, which is more efficient than exact methods especially on large graphs with many triangles.

Contribution

It presents the first efficient approximation algorithms for graph trussness with provable guarantees, reducing computational complexity compared to exact methods.

Findings

The approximation algorithm is faster than exact methods on graphs with many triangles.

It achieves a (1 ± ε)-approximation with lower asymptotic complexity.

Exact computation complexity is shown to be inherently tied to triangle counting complexity.

Abstract

A $k$-truss is an edge-induced subgraph $H$ such that each of its edges belongs to at least $k-2$ triangles of $H$. This notion has been introduced around ten years ago in social network analysis and security, as a form of cohesive subgraph that is rich of triangles and less stringent than the clique. The \emph{trussness} of a graph is the maximum $k$ such that a $k$-truss exists. The problem of computing $k$-trusses has been largely investigated from the practical and engineering point of view. On the other hand, the theoretical side of the problem has received much less attention, despite presenting interesting challenges. The existing methods share a common design, based on iteratively removing the edge with smallest support, where the support of an edge is the number of triangles containing it. The aim of this paper is studying algorithmic aspects of graph trussness. While it is possible to show that the time complexity of computing exactly the graph trussness and that of counting/listing all triangles is inherently the same, we provide efficient algorithms for estimating its value, under suitable conditions, with significantly lower complexity than the exact approach. In particular, we provide a $(1 \pm \epsilon)$-approximation algorithm that is asymptotically faster than the exact approach, on graphs which contain $\omega(m \, polylog(n))$ triangles, and has the same running time on graphs that do not. For the latter case, we also show that it is impossible to obtain an approximation algorithm with faster running time than the one of the exact approach when the number of triangles is $O(m)$, unless well known conjectures on triangle-freeness and Boolean matrix multiplication are false.