On the Generalized Mean Densest Subgraph Problem: Complexity and Algorithms

TL;DR

This paper investigates the computational complexity of the generalized p-mean densest subgraph problem, proving NP-hardness for certain p ranges, and introduces efficient approximation algorithms with strong empirical performance.

Contribution

It establishes NP-hardness for specific p values, provides approximation algorithms for all p<1, and develops a fast implementation for p>1, advancing understanding and solution methods for the problem.

Findings

NP-hardness for p in (-1/8, 0) and (0, 1/4)

Two 1/2-approximation algorithms for p<1

Fast near-linear implementation for p>1

Abstract

Dense subgraph discovery is an important problem in graph mining and network analysis with several applications. Two canonical problems here are to find a maxcore (subgraph of maximum min degree) and to find a densest subgraph (subgraph of maximum average degree). Both of these problems can be solved in polynomial time. Veldt, Benson, and Kleinberg [VBK21] introduced the generalized $p$-mean densest subgraph problem which captures the maxcore problem when $p=-\infty$ and the densest subgraph problem when $p=1$. They observed that the objective leads to a supermodular function when $p \ge 1$ and hence can be solved in polynomial time; for this case, they also developed a simple greedy peeling algorithm with a bounded approximation ratio. In this paper, we make several contributions. First, we prove that for any $p \in (-\frac{1}{8}, 0) \cup (0, \frac{1}{4})$ the problem is NP-Hard and for any $p \in (-3,0) \cup (0,1)$ the weighted version of the problem is NP-Hard, partly resolving a question left open in [VBK21]. Second, we describe two simple $1/2$-approximation algorithms for all $p < 1$, and show that our analysis of these algorithms is tight. For $p > 1$ we develop a fast near-linear time implementation of the greedy peeling algorithm from [VBK21]. This allows us to plug it into the iterative peeling algorithm that was shown to converge to an optimum solution [CQT22]. We demonstrate the efficacy of our algorithms by running extensive experiments on large graphs. Together, our results provide a comprehensive understanding of the complexity of the $p$-mean densest subgraph problem and lead to fast and provably good algorithms for the full range of $p$.