Improved Competitive Ratios for Online Bipartite Matching on Degree Bounded Graphs

TL;DR

This paper improves the competitive ratios for online bipartite matching on degree-bounded graphs using randomized algorithms, addressing open questions and extending results to stochastic models and b-matching scenarios.

Contribution

It introduces new randomized algorithms with better competitive ratios for adversarial and stochastic models, solving open problems from prior work on degree-bounded graphs.

Findings

Achieved a competitive ratio of 1 - (1-1/d)^k + Ω(d^{-4}·e^{-k/d}) for adversarial arrivals.

Established a minimum competitive ratio of 0.8237 in the stochastic model.

Provided lower bounds for the b-matching problem in both models.

Abstract

We consider the online bipartite matching problem on $(k,d)$-bounded graphs, where each online vertex has at most $d$ neighbors, each offline vertex has at least $k$ neighbors, and $k\geq d\geq 2$. The model of $(k,d)$-bounded graphs is proposed by Naor and Wajc (EC 2015 and TEAC 2018) to model the online advertising applications in which offline advertisers are interested in a large number of ad slots, while each online ad slot is interesting to a small number of advertisers. They proposed deterministic and randomized algorithms with a competitive ratio of $1 - (1-1/d)^k$ for the problem, and show that the competitive ratio is optimal for deterministic algorithms. They also raised the open questions of whether strictly better competitive ratios can be achieved using randomized algorithms, for both the adversarial and stochastic arrival models. In this paper we answer both of their open problems affirmatively. For the adversarial arrival model, we propose a randomized algorithm with competitive ratio $1 - (1-1/d)^k + \Omega(d^{-4}\cdot e^{-\frac{k}{d}})$ for all $k\geq d\geq 2$. We also consider the stochastic model and show that even better competitive ratios can be achieved. We show that for all $k\geq d\geq 2$, the competitive ratio is always at least $0.8237$. We further consider the $b$-matching problem when each offline vertex can be matched at most $b$ times, and provide several competitive ratio lower bounds for the adversarial and stochastic model.