K-Means Clustering

A basic Javascript implementation of the [cluster analysis] 1 algorithm.

Usage

• 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
    }
];


var ranges = require('dataset').findRanges(params, data);
var normalized = require('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;

var means = require('kmeans').algorithm(points, k, console.log);

The call to algorithm() 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.

var pbk = require('phamBestK');

var maxKToTest = 10;
var result = pbk.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));
}

Todo

• denormalize data
• provide ability to label data points, dimensions and means
• build an asynchronous version of the algorithm