English

An adaptive multiclass nearest neighbor classifier

Machine Learning 2019-11-05 v4 Machine Learning

Abstract

We consider a problem of multiclass classification, where the training sample Sn={(Xi,Yi)}i=1nS_n = \{(X_i, Y_i)\}_{i=1}^n is generated from the model P(Y=mX=x)=ηm(x)\mathbb P(Y = m | X = x) = \eta_m(x), 1mM1 \leq m \leq M, and η1(x),,ηM(x)\eta_1(x), \dots, \eta_M(x) are unknown α\alpha-Holder continuous functions.Given a test point XX, our goal is to predict its label. A widely used k\mathsf k-nearest-neighbors classifier constructs estimates of η1(X),,ηM(X)\eta_1(X), \dots, \eta_M(X) and uses a plug-in rule for the prediction. However, it requires a proper choice of the smoothing parameter k\mathsf k, which may become tricky in some situations. In our solution, we fix several integers n1,,nKn_1, \dots, n_K, compute corresponding nkn_k-nearest-neighbor estimates for each mm and each nkn_k and apply an aggregation procedure. We study an algorithm, which constructs a convex combination of these estimates such that the aggregated estimate behaves approximately as well as an oracle choice. We also provide a non-asymptotic analysis of the procedure, prove its adaptation to the unknown smoothness parameter α\alpha and to the margin and establish rates of convergence under mild assumptions.

Keywords

Cite

@article{arxiv.1804.02756,
  title  = {An adaptive multiclass nearest neighbor classifier},
  author = {Nikita Puchkin and Vladimir Spokoiny},
  journal= {arXiv preprint arXiv:1804.02756},
  year   = {2019}
}

Comments

Accepted in ESAIM: Probability & Statistics. The original publication is available at www.esaim-ps.org

R2 v1 2026-06-23T01:17:25.589Z