1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157


title: Detecting exact nearest neighbors
author:
 name: Aaron Lun
affiliation: Cancer Research UK Cambridge Institute, Cambridge, United Kingdom
date: "Revised: 2 December 2018"
output:
BiocStyle::html_document:
toc_float: true
package: BiocNeighbors
vignette: >
%\VignetteIndexEntry{1. Detecting exact nearest neighbors}
%\VignetteEngine{knitr::rmarkdown}
%\VignetteEncoding{UTF8}
bibliography: ref.bib

```{r, echo=FALSE, results="hide", message=FALSE}
require(knitr)
opts_chunk$set(error=FALSE, message=FALSE, warning=FALSE)
library(BiocNeighbors)
```
# Introduction
The `r Biocpkg("BiocNeighbors")` package implements a few algorithms for exact nearest neighbor searching:
 The kmeans for knearest neighbors (KMKNN) algorithm [@wang2012fast] uses kmeans clustering to create an index.
Within each cluster, the distance of each of that cluster's points to the cluster center are computed and used to sort all points.
Given a query point, the distance to each cluster center is determined and the triangle inequality is applied to determine which points in each cluster warrant a full distance calculation.
 The vantage point (VP) tree algorithm [@yianilos1993data] involves constructing a tree where each node is located at a data point and is associated with a subset of neighboring points.
Each node progressively partitions points into two subsets that are either closer or further to the node than a given threshold.
Given a query point, the triangle inequality is applied at each node in the tree to determine if the child nodes warrant searching.
 The exhaustive search is a simple bruteforce algorithm that computes distances to between all data and query points.
This has the worst computational complexity but can actually be faster than the other exact algorithms in situations where indexing provides little benefit, e.g., data sets with few points and/or a very large number of dimensions.
Both KMKNN and VPtrees involve a component of randomness during index construction, though the knearest neighbors result is fully deterministic^[Except in the presence of ties, see `?"BiocNeighborsties"` for details.].
# Identifying knearest neighbors
The most obvious application is to perform a knearest neighbors search.
We'll mock up an example here with a hypercube of points, for which we want to identify the 10 nearest neighbors for each point.
```{r}
nobs < 10000
ndim < 20
data < matrix(runif(nobs*ndim), ncol=ndim)
```
The `findKNN()` method expects a numeric matrix as input with data points as the rows and variables/dimensions as the columns.
We indicate that we want to use the KMKNN algorithm by setting `BNPARAM=KmknnParam()` (which is also the default, so this is not strictly necessary here).
We could use a VP tree instead by setting `BNPARAM=VptreeParam()`.
```{r}
fout < findKNN(data, k=10, BNPARAM=KmknnParam())
head(fout$index)
head(fout$distance)
```
Each row of the `index` matrix corresponds to a point in `data` and contains the row indices in `data` that are its nearest neighbors.
For example, the 3rd point in `data` has the following nearest neighbors:
```{r}
fout$index[3,]
```
... with the following distances to those neighbors:
```{r}
fout$distance[3,]
```
Note that the reported neighbors are sorted by distance.
# Querying knearest neighbors
Another application is to identify the knearest neighbors in one dataset based on query points in another dataset.
Again, we mock up a small data set:
```{r}
nquery < 1000
ndim < 20
query < matrix(runif(nquery*ndim), ncol=ndim)
```
We then use the `queryKNN()` function to identify the 5 nearest neighbors in `data` for each point in `query`.
```{r}
qout < queryKNN(data, query, k=5, BNPARAM=KmknnParam())
head(qout$index)
head(qout$distance)
```
Each row of the `index` matrix contains the row indices in `data` that are the nearest neighbors of a point in `query`.
For example, the 3rd point in `query` has the following nearest neighbors in `data`:
```{r}
qout$index[3,]
```
... with the following distances to those neighbors:
```{r}
qout$distance[3,]
```
Again, the reported neighbors are sorted by distance.
# Further options
Users can perform the search for a subset of query points using the `subset=` argument.
This yields the same result as but is more efficient than performing the search for all points and subsetting the output.
```{r}
findKNN(data, k=5, subset=3:5)
```
If only the indices are of interest, users can set `get.distance=FALSE` to avoid returning the matrix of distances.
This will save some time and memory.
```{r}
names(findKNN(data, k=2, get.distance=FALSE))
```
It is also simple to speed up functions by parallelizing the calculations with the `r Biocpkg("BiocParallel")` framework.
```{r}
library(BiocParallel)
out < findKNN(data, k=10, BPPARAM=MulticoreParam(3))
```
For multiple queries to a constant `data`, the preclustering can be performed in a separate step with `buildIndex()`.
The result can then be passed to multiple calls, avoiding the overhead of repeated clustering^[The algorithm type is automatically determined when `BNINDEX` is specified, so there is no need to also specify `BNPARAM` in the later functions.].
```{r}
pre < buildIndex(data, BNPARAM=KmknnParam())
out1 < findKNN(BNINDEX=pre, k=5)
out2 < queryKNN(BNINDEX=pre, query=query, k=2)
```
The default setting is to search on the Euclidean distance.
Alternatively, we can use the Manhattan distance by setting `distance="Manhattan"` in the `BiocNeighborParam` object.
```{r}
out.m < findKNN(data, k=5, BNPARAM=KmknnParam(distance="Manhattan"))
```
Advanced users may also be interested in the `raw.index=` argument, which returns indices directly to the precomputed object rather than to `data`.
This may be useful inside package functions where it may be more convenient to work on a common precomputed object.
# Session information
```{r}
sessionInfo()
```
# References
