PegasosSVM

A simple, lightweight package for linear classification based on the support vector machine as described in this study.

Creating a model

const PegasosSVM = require('pegasossvm');
 
const inputValues = [[...], [...], ...]; // input array with each item being its own n-dimensional feature array
const outputValues = [...]; // these do NOT need to be mapped to -1 and 1, that will be done internally
 
const options = { // the completely optional additional parameters, shown here at their defaults
  lambda: 1, // regularization parameter
  k: 1, // change if you want to use mini-batch learning in each iteration
  weights: [0, ..., 0], // initial weights, start at 0 by default
  projection: true, // use a projection step; this was shown to have little effect but is offered anyway
}
 
const model = new PegasosSVM(inputValues, outputValues, options);
 
model.train(iterations);
console.log(model.predict(testValue)); //=> predicted output

Notes

• As the algorithm is intended to decrease the learning rate as the training progresses, a regularization parameter higher than other Perceptrons' learning rate is often helpful.
• The algorithm selects random batches of input/output pairs to train on in each iteration, and if it selects unrepresentative pairs within the first few iterations (when the learning rate is still high) then it will have very poor performance.
• If k = 1, then the special case algorithm is used, otherwise the general algorithm for a mini-batch of size k is used.
• The train method returns its instance, so method calls can be chained.