Reproduce “Extrapolation Towards Imaginary $0$-Nearest Neighbour and Its Improved Convergence Rate” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Akifumi Okuno, Hidetoshi Shimodaira
The paper introduces Multiscale k-Nearest Neighbour (MS-k-NN), a classifier that eradicates asymptotic bias by extrapolating standard k-NN estimators towards an "imaginary 0-NN" point using polynomial regression. This approach achieves the optimal convergence rate for excess risk while implicitly deriving adaptive real-valued weights, demonstrated by superior performance on diverse datasets.
Key insights
- Asymptotic bias in k-NN can be effectively eradicated by extrapolating the relationship between k-NN estimates and local neighborhood radii towards a hypothetical "0-Nearest Neighbour" point.
- The proposed multiscale regression implicitly generates optimal, adaptive real-valued weights for k-NN, which are crucial for achieving state-of-the-art convergence rates without explicit, complex calculations.
- The method bridges the gap between the simplicity of k-NN and the theoretical optimality of advanced non-parametric classifiers, making high-performance k-NN practically accessible.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.