English

Efficient Linear Bandits through Matrix Sketching

Machine Learning 2022-03-22 v3 Machine Learning

Abstract

We prove that two popular linear contextual bandit algorithms, OFUL and Thompson Sampling, can be made efficient using Frequent Directions, a deterministic online sketching technique. More precisely, we show that a sketch of size mm allows a O(md)\mathcal{O}(md) update time for both algorithms, as opposed to Ω(d2)\Omega(d^2) required by their non-sketched versions in general (where dd is the dimension of context vectors). This computational speedup is accompanied by regret bounds of order (1+εm)3/2dT(1+\varepsilon_m)^{3/2}d\sqrt{T} for OFUL and of order ((1+εm)d)3/2T\big((1+\varepsilon_m)d\big)^{3/2}\sqrt{T} for Thompson Sampling, where εm\varepsilon_m is bounded by the sum of the tail eigenvalues not covered by the sketch. In particular, when the selected contexts span a subspace of dimension at most mm, our algorithms have a regret bound matching that of their slower, non-sketched counterparts. Experiments on real-world datasets corroborate our theoretical results.

Keywords

Cite

@article{arxiv.1809.11033,
  title  = {Efficient Linear Bandits through Matrix Sketching},
  author = {Ilja Kuzborskij and Leonardo Cella and Nicolò Cesa-Bianchi},
  journal= {arXiv preprint arXiv:1809.11033},
  year   = {2022}
}
R2 v1 2026-06-23T04:22:03.587Z