Percipio - Easy Data Science (Machine Learning) in JavaScript & Node

Percipio is a simple minimalistic JavaScript library for understanding & making decisions with data.

Features

• Bayesian Bandit algorithm (using Thompson sampling)
• Naive Bayes classifier

Install

npm install percipio

Quick Start

Let's find out which programming language is better! Java or C#, anyone? (this might be a bit contrived example...) We can model this using simple Multi-armed bandit experiment (Multi-armed bandit experiments are even used by Google)

Experiment setup

We define 2 arms (possible outcomes) as follows

• <em>Arm 1</em> - id: 1, reward: Java
• <em>Arm 2</em> - id: 2, reward: C#

and create the Bandit predictor

var bandits = require('percipio').bandits
var BanditPredictor = bandits.Predictor

var rewards = ["Java", "C#"]
var armIds = [0, 1]

var predictor = BanditPredictor([
    bandits.createArm(armIds[0], rewards[0]),
    bandits.createArm(armIds[1], rewards[1])
])

Hidden probabilities

Next let's choose the probabilities which the predictor should find

var hiddenProbabilities = [0.5, 0.7] 

Simulation

Let's define our result simulation function (in the real world you should get results from your app, users etc.)

function simulateResult(p){
    return Math.random() < p ? 1 : 0
}

And run the simulation

for (var i = 0; i < 1000; i++) {
    var arm = predictor.predict() 
    var p = hiddenProbabilities[arm.id]
    predictor.learn(arm, simulateResult(p))
}

Result

Now the predictor has (hopefully) learned the hidden probabilities and we can get them

var javaProbabilities = predictor.posteriorProbabilities()[0]
var cSharpProbabilities = predictor.posteriorProbabilities()[1]
console.log(javaProbabilities)
console.log(cSharpProbabilities)

Complete example

Now try to run this yourself

var bandits = require('percipio').bandits
var BanditPredictor = bandits.Predictor

var rewards = ["Java", "C#"]
var armIds = [0, 1]

var predictor = BanditPredictor([
    bandits.createArm(armIds[0], rewards[0]),
    bandits.createArm(armIds[1], rewards[1])
])

var hiddenProbabilities = [0.5, 0.7]

function simulateResult(p){
    return Math.random() < p ? 1 : 0
}

for (var i = 0; i < 1000; i++) {
    var arm = predictor.predict() 
    var p = hiddenProbabilities[arm.id]
    predictor.learn(arm, simulateResult(p))
}

var javaProbabilities = predictor.posteriorProbabilities()[0]
var cSharpProbabilities = predictor.posteriorProbabilities()[1]
console.log(javaProbabilities)
console.log(cSharpProbabilities)

Current state

Pretty alphaish, I guess. Looking forward to implement

• kNN
• Linear regression
• Data loaders/importers

Wanna help out?

Hop right in!

Development setup

git clone git@github.com:naughtyspirit/percipio.git
cd percipio
npm install

Run tests

npm test

License

MIT