Some Improved Results on Fair and Balanced Graph Partitions

TL;DR

This paper studies fair graph partitioning, providing improved guarantees for envy-freeness and core stability, and offers efficient algorithms under relaxed constraints, addressing open questions in the field.

Contribution

It introduces new approximate fairness and stability guarantees for balanced graph partitions and develops efficient algorithms for their computation.

Findings

Existence of balanced partitions that are approximately envy-free and in the core.

These guarantees are comparable or better than previous bounds.

Efficient algorithms are provided when relaxing the balancedness constraint.

Abstract

We consider the problem of partitioning an undirected graph (representing a social network) over $n$ nodes and max degree $\Delta$ into $k$ equally sized parts. Each node in the graph, representing an agent, derives utility proportional to the number of their neighbors in their assigned part. Our goal is to find a balanced partitioning that is fair. The two notions of fairness we consider are the core and envy-freeness. A partition is envy-free if no node gains utility from moving to a different part, and a partition is in the core if no set of $n/k$ nodes can deviate to form a new part with all nodes gaining in utility. We show that there exists a balanced partition which is both $O(\max\{\sqrt{\Delta}, k^2\} \ln n)$-approximately envy-free and in the $(k + o(k))$-approximate core. Taken separately, these two guarantees are comparable to (and in some cases, better than) the best known envy-freeness and core guarantees for this problem. Moreover, we show that these desirable partitions can be computed efficiently if we slightly relax the balancedness constraint. In addition, when $k = 2$, we show that a $(1.618 + o(1))$-core exists, and a $(2 + \varepsilon)$-core can be computed in polynomial time. The last two results make progress on two open questions from Li et al. [AAAI, 2023].