Sushma Kumari
(Kyoto university)
The $k$-nearest neighbour (NN) rule is one of the important learning rules in machine learning. We start with the definition of the $k$-NN rule and universal consistency. Charles Stone proved the universal consistency of $k$-NN rule in $\mathbb{R}^d$. The essential ingredient for Stone’s theorem was the so-called geometric Stone’s lemma but because of the structure of $\mathbb{R}^d$, the proof of the geometric Stone’s lemma is restricted to finite dimensional normed spaces only. Cerou and Guyader in 2006, proved that the $k$-NN rule is universal consistent in more general metric spaces which satisfy the Lebesgue-Besicovitch differentiation theorem. Further, David Preiss proved that the metric spaces satisfying the Lebesgue-Besicovitch differentiation theorem are the metric spaces with the metric having $\sigma$-finite dimension (called as $\sigma$-finite dimensional metric spaces).
In this work, we try to extend the geometric Stone’s lemma to the $\sigma$-finite dimensional metric spaces and reprove the universal consistency using the generalized Stone’s theorem. Further questions related to this work are also discussed.