Open Source: JavaScript implementation of Linear and Kernel SVM using Gradient Descent

js-svm

This package provides javascript implementation of linear SVM and SVM with gaussian kernel:

https://github.com/chen0040/js-svm

Features

• Support for binary classification
• Support for multi-class classification

Install

npm install js-svm

Usage

SVM Binary Classifier

The sample code below show how to use SVM binary classifier on the iris datsets to classify whether a data row belong to species Iris-virginica:

var jssvm = require('js-svm');
var iris = require('js-datasets-iris');

var svm = new jssvm.BinarySvmClassifier();

iris.shuffle();

var trainingDataSize = Math.round(iris.rowCount * 0.9);
var trainingData = [];
var testingData = [];
for(var i=0; i < iris.rowCount ; ++i) {
   var row = [];
   row.push(iris.data[i][0]); // sepalLength;
   row.push(iris.data[i][1]); // sepalWidth;
   row.push(iris.data[i][2]); // petalLength;
   row.push(iris.data[i][3]); // petalWidth;
   row.push(iris.data[i][4] == "Iris-virginica" ? 1.0 : 0.0); // output which is 1 if species is Iris-virginica; 0 otherwise
   if(i < trainingDataSize){
        trainingData.push(row);
   } else {
       testingData.push(row);
   }
}


var result = svm.fit(trainingData);

console.log(result);

for(var i=0; i < testingData.length; ++i){
   var predicted = svm.transform(testingData[i]);
   console.log("actual: " + testingData[i][4] + " predicted: " + predicted);
}

To configure the BinarySvmClassifier, use the following code when it is created:

var svm = new jssvm.BinarySvmClassifier({
   alpha: 0.01, // learning rate
   iterations: 1000, // maximum iterations
   C: 5.0, // panelty term
   trace: false // debug tracing
});

Multi-Class Classification using One-vs-All Logistic Regression

The sample code below illustrates how to run the multi-class classifier on the iris datasets to classifiy the species of each data row:

var jssvm = require('js-svm');
var iris = require('js-datasets-iris');

var classifier = new jssvm.MultiClassSvmClassifier();

iris.shuffle();

var trainingDataSize = Math.round(iris.rowCount * 0.9);
var trainingData = [];
var testingData = [];
for(var i=0; i < iris.rowCount ; ++i) {
   var row = [];
   row.push(iris.data[i][0]); // sepalLength;
   row.push(iris.data[i][1]); // sepalWidth;
   row.push(iris.data[i][2]); // petalLength;
   row.push(iris.data[i][3]); // petalWidth;
   row.push(iris.data[i][4]); // output is species
   if(i < trainingDataSize){
        trainingData.push(row);
   } else {
       testingData.push(row);
   }
}


var result = classifier.fit(trainingData);

console.log(result);

for(var i=0; i < testingData.length; ++i){
   var predicted = classifier.transform(testingData[i]);
   console.log("svm prediction testing: actual: " + testingData[i][4] + " predicted: " + predicted);
}

To configure the MultiClassSvmClassifier, use the following code when it is created:

var classifier = new jssvm.MultiClassSvmClassifier({
   alpha: 0.01, // learning rate
   iterations: 1000, // maximum iterations
   C: 5.0 // panelty term
   sigma: 1.0 // the standard deviation for the gaussian kernel
});

Switch between linear and guassian kernel

By default the kernel used by the binary and multi-class classifier is "linear" which can be printed by:

console.log(classifier.kernel);

To switch to use gaussian kernel, put the property 'kernel: "gaussian"' in the config data when the classifier is created:

var svm = new jssvm.BinarySvmClassifier({
   ...,
   kernel: 'gaussian'
});

....

var svm = new jssvm.MultiClassSvmClassifier({
   ...,
   kernel: 'gaussian'
});

Usage In HTML

Include the "node_modules/js-svm/build/jssvm.min.js" (or "node_modules/js-svm/src/jssvm.js") in your HTML <script> tag

The code in the script tag looks sth like this:

var classifier = new jssvm.BinarySvmClassifier();