A basic Javascript implementation of the [cluster analysis] 1 algorithm.
npm i @jbeuckm/k-means-js --save
- Optionally, normalize the data.
The normalizer will scale numerical data between [0,1] and will generate n outputs of either zero or one for discrete data, eg. category.
// Tell the normalizer about the category field.
var params = {
category: "discrete"
};
// Category is a discrete field with two possible values.
// Value is a linear field with continuous possible values.
var data = [
{
category: "a",
value: 25
},
{
category: "b",
value: 7.6
},
{
category: "a",
value: 28
}
];
import { dataset } from '@jbeuckm/k-means-js'
var ranges = dataset.findRanges(params, data);
var normalized = dataset.normalize(data, ranges);- Run the algorithm.
// This non-normalized sample data with n=k is a pretty awful example.
var points = [
[.1, .2, .3],
[.4, .5, .6],
[.7, .8, .9]
];
var k = 3;
import kmeans from '@jbeuckm/k-means-js'
var means = kmeans.cluster(points, k, console.log);The call to cluster() will find the data's range in each dimension, generate k=3 random points, and iterate until the means are static.
- Find the best K
The method described by Pham, et al. is implemented. The algorithm evaluates K-means repeatedly for different values of K, and returns the best (guess) value for K as well as the set of means found during evaluation.
import { phamBestK } from '@jbeuckm/k-means-js'
var maxKToTest = 10;
var result = phamBestK.findBestK(points, maxKToTest);
console.log("this data has "+result.K+" clusters");
console.log("cluster centroids = "+result.means);- Denormalize data
Denormalization can be used to show the means discovered:
for (var i= 0, l=result.means.length; i<l; i++) {
console.log(dataset.denormalizeDatum(result.means[i], ranges));
}- denormalize data
- provide ability to label data points, dimensions and means
- build an asynchronous version of the algorithm