Chromatic $k$-Nearest Neighbor Queries

TL;DR

This paper introduces efficient data structures for chromatic k-nearest neighbor queries that identify the most frequent color among the nearest points, achieving query times independent of k with near-linear space, and also explores approximate solutions.

Contribution

The paper presents novel data structures for chromatic k-NN queries with query times independent of k and near-linear space, including approximate methods for faster responses.

Findings

Linear space data structures with $O(n^{1/2} ext{ log } n)$ query time in $ ext{R}^1$

Query times between $O(n^{2/3} ext{ log}^{2/3} n)$ and $O(n^{5/6} ext{ polylog} n)$ in $ ext{R}^2$

Approximate solutions with faster query times depending on $ ext{ε}$ in $ ext{R}^1$ and $ ext{R}^2$

Abstract

Let $P$ be a set of $n$ colored points. We develop efficient data structures that store $P$ and can answer chromatic $k$-nearest neighbor ($k$-NN) queries. Such a query consists of a query point $q$ and a number $k$, and asks for the color that appears most frequently among the $k$ points in $P$ closest to $q$. Answering such queries efficiently is the key to obtain fast $k$-NN classifiers. Our main aim is to obtain query times that are independent of $k$ while using near-linear space. We show that this is possible using a combination of two data structures. The first data structure allow us to compute a region containing exactly the $k$-nearest neighbors of a query point $q$, and the second data structure can then report the most frequent color in such a region. This leads to linear space data structures with query times of $O(n^{1 / 2} \log n)$ for points in $\mathbb{R}^1$, and with query times varying between $O(n^{2/3}\log^{2/3} n)$ and $O(n^{5/6} {\rm polylog} n)$, depending on the distance measure used, for points in $\mathbb{R}^2$. Since these query times are still fairly large we also consider approximations. If we are allowed to report a color that appears at least $(1-\varepsilon)f^*$ times, where $f^*$ is the frequency of the most frequent color, we obtain a query time of $O(\log n + \log\log_{\frac{1}{1-\varepsilon}} n)$ in $\mathbb{R}^1$ and expected query times ranging between $\tilde{O}(n^{1/2}\varepsilon^{-3/2})$ and $\tilde{O}(n^{1/2}\varepsilon^{-5/2})$ in $\mathbb{R}^2$ using near-linear space (ignoring polylogarithmic factors).