kNN.jl

Note: This library is not activly developed. See NearestNeighbors.jl for an updated alternative.

Basic k-nearest neighbors classification and regression.

For a list of the distance metrics that can be used in k-NN classification, see Distances.jl

For a list of the smoothing kernels that can be used in kernel regression, see SmoothingKernel.jl

Example 1: Classification using Euclidean distance

using kNN
using DataArrays
using DataFrames
using RDatasets
using Distances

iris = dataset("datasets", "iris")
X = array(iris[:, 1:4])'
y = array(iris[:, 5])
model = knn(X, y, metric = Euclidean())

predict_k1 = predict(model, X, 1)
predict_k2 = predict(model, X, 2)
predict_k3 = predict(model, X, 3)
predict_k4 = predict(model, X, 4)
predict_k5 = predict(model, X, 5)

mean(predict_k1 .== y)
mean(predict_k2 .== y)
mean(predict_k3 .== y)
mean(predict_k4 .== y)
mean(predict_k5 .== y)

Example 2: Classification using Manhattan distance

using kNN
using DataArrays
using DataFrames
using RDatasets
using Distances

iris = dataset("datasets", "iris")
X = array(iris[:, 1:4])'
y = array(iris[:, 5])
model = knn(X, y, metric = Cityblock())

predict_k1 = predict(model, X, 1)
predict_k2 = predict(model, X, 2)
predict_k3 = predict(model, X, 3)
predict_k4 = predict(model, X, 4)
predict_k5 = predict(model, X, 5)

mean(predict_k1 .== y)
mean(predict_k2 .== y)
mean(predict_k3 .== y)
mean(predict_k4 .== y)
mean(predict_k5 .== y)

Example 3: Regression using Gaussian kernel

using Base.Test
using kNN
using StatsBase

srand(1)
n = 1_000
x = 10 * randn(n)
y = sin(x) + 0.5 * randn(n)

fit = kernelregression(x, y, kernel = :gaussian)
grid = minimum(x):0.1:maximum(x)
predictions = predict(fit, grid)

To Do

Allow user to request that knn generate a ball tree, KD-tree or cover tree as a method for conducting nearest neighbor searches.
Allow user to request that approximate nearest neighbors be returned instead of exact nearest neighbors.
Clean up API