Bayes Text Classifier

Install

npm i bayes-text-classifier

Usage

import { BayesTextClassifier } from 'bayes-text-classifier';

const classifier = new BayesTextClassifier();
classifier.learnText("i am feeling good today", "good");
classifier.learnText("i am feeling bad today", "bad");
console.log(classifier.classifyText("im good"));
// good
console.log(classifier.classifyText("im bad"));
// bad

How It Works

This is a an implementation of a multinomial naive bayes classifier in pure javascript, which means you can use it anywhere you can use javascript.

What is a Naive Bayes Classifier?

Naive bayes classifiers are simple classifiers often used in text classification tasks, such as, famously, detecting whether or not a given email is spam or not spam. Naive bayes classifiers are useful when you have sparse data, intractably many parameters, or when you to train the model in linear time.

https://en.wikipedia.org/wiki/Naive_Bayes_classifier

Examples

Persistence

You can persist your model to a string using the dump method.

const classifier = new BayesTextClassifier();
classifier.learnText("i am feeling good today", "good");
classifier.learnText("i am feeling bad today", "bad");
const str = classifier.dump();
console.log(str);
// {"tokenizePunctuation":true,"multinomialNaiveBayesClassifier":
// "{\"alpha\":1e-9,\"classRelativeFeatureFrequencies\"
// :[[\"good\",\"[[\\\"i\\\",1],[\\\"am\\\",1],[\\\"feeling\\\",1],
// [\\\"good\\\",1],[\\\"today\\\",1]]\"],[\"bad\",\"[[\\\"i\\\",1],
// [\\\"am\\\",1],[\\\"feeling\\\",1],[\\\"bad\\\",1],[\\\"today\\\",
// 1]]\"]],\"classTypeFrequencyCounter\":\"[[\\\"good\\\",1],
// [\\\"bad\\\",1]]\",\"n\":2}"}

Reloading the model is as easy as:

const newClassifier = BayesTextClassifier.load<string>(str);
console.log(newClassifier.classifyText("im good"));
// good
console.log(newClassifier.classifyText("im bad"));
// bad

Classifying Spam vs "Ham"

In this classic example, we load a dataset of spam vs. not-spam ("ham") emails and attempt to classify each email correctly. We split the data into randomized training and testing sets, and then just learn on each example.

const fs = require("fs");
const { BayesTextClassifier } = require("bayes-text-classifier");
const trainTestSplit = require("train-test-split");

const data = fs
  .readFileSync("./spam_ham.txt")
  .toString()
  .split("\n")
  .map((s) => s.split("\t"));
const clf = new BayesTextClassifier();
const [train, test] = trainTestSplit(data, 0.8);
for (const [category, text] of train) {
  clf.learnText(text, category);
}
let correct = 0;
for (const [category, text] of test) {
  if (clf.classifyText(text) === category) {
    ++correct;
  }
}
console.log("Accuracy=", correct / test.length || 0);

We can see that this model gives great performance. Since the model is validated on a random subset of the data, I've ran the above code 5 times, so you can see its output on 5 random subsets of the data. An accuracy above 90% can be considered great performance.

➜  examples git:(master) node spam-classifier.js
Accuracy= 0.9775784753363229
➜  examples git:(master) node spam-classifier.js
Accuracy= 0.9775784753363229
➜  examples git:(master) node spam-classifier.js
Accuracy= 0.9847533632286996
➜  examples git:(master) node spam-classifier.js
Accuracy= 0.9820627802690582
➜  examples git:(master) node spam-classifier.js
Accuracy= 0.9847533632286996

Please note that the above dataset may or may not be as challenging as a real life dataset. The point of this example is to show that the model works, and is simple to use.