Universal consistency of the $k$-NN rule in metric spaces and Nagata dimension. III
A new arXiv paper establishes a crucial link between the k-nearest neighbor (k-NN) classification rule and the Nagata dimension of metric spaces. The research demonstrates that the k-NN classifier is universally consistent if and only if the space possesses the strong Lebesgue-Besicovitch differentiation property or is sigma-finite dimensional. The paper also clarifies that the weak Lebesgue-Besicovitch property is insufficient for k-NN consistency, even providing a counterexample on the real line with a modified metric. AI
IMPACT Establishes theoretical underpinnings for k-NN classifier performance in complex metric spaces, potentially guiding future algorithm development.