Bernoulli Naive Bayes

Algorithm overview

Bernoulli Naive Bayes is a probabilistic classifier designed for binary-valued features. Instead of modeling continuous numeric values directly, it focuses on whether a feature is present or absent, active or inactive, true or false.

That makes it a strong option when:

• your input features are boolean indicators, flags, or token-presence signals
• you need a lightweight baseline for text, events, or sparse binary data
• you want class probabilities from a simple model that is easy to run in production

It is often used for fast classification tasks such as keyword detection, binary event modeling, and simple document labeling.

JavaScript implementation

@kanaries/ml exposes Bernoulli Naive Bayes as a pure JavaScript and TypeScript estimator, which is useful when your feature extraction already happens in the browser or in a Node.js request pipeline. You can keep boolean features, keyword flags, or binary event signals inside the same JS application without handing them off to a separate Python service.

That makes this implementation a good fit for lightweight text classification, rule-assisted scoring, and product features where simple probabilistic models need to run close to the UI or application logic.

Quick start

from sklearn.naive_bayes import BernoulliNB

X = [[1, 0, 1], [1, 1, 0], [0, 1, 1], [0, 0, 1]]
y = [1, 1, 0, 0]

clf = BernoulliNB(alpha=1.0)
clf.fit(X, y)
pred = clf.predict([[1, 0, 0], [0, 1, 1]])

import { Bayes } from '@kanaries/ml';

const X = [[1, 0, 1], [1, 1, 0], [0, 1, 1], [0, 0, 1]];
const y = [1, 1, 0, 0];

const clf = new Bayes.BernoulliNB({ alpha: 1.0 });
clf.fit(X, y);
const pred = clf.predict([[1, 0, 0], [0, 1, 1]]);

Quick JavaScript example

import { Bayes } from '@kanaries/ml';

const trainX = [[1, 0, 1], [1, 1, 0], [0, 1, 1], [0, 0, 1]];
const trainY = [1, 1, 0, 0];

const model = new Bayes.BernoulliNB({
    alpha: 1.0,
    binarize: null,
});

model.fit(trainX, trainY);

const predictions = model.predict([[1, 0, 0], [0, 1, 1]]);
console.log(predictions);

Detailed API reference

Bernoulli Naive Bayes is designed for binary or boolean features. Continuous features can be converted into binary values using the binarize threshold, and class probabilities are estimated with additive smoothing.

interface BernoulliNBProps {
    alpha?: number;
    binarize?: number | null;
    fitPrior?: boolean;
    classPrior?: number[] | null;
}
constructor(props: BernoulliNBProps = {})

Parameters

• alpha — Additive smoothing parameter applied when estimating probabilities. Defaults to 1.0.
• binarize — Threshold for binarizing input features. If null, the input is assumed to already be binary.
• fitPrior — Whether to learn class prior probabilities from the training data. When false, a uniform prior is used.
• classPrior — Optional array of prior probabilities for each class. If provided, these values override the learned priors.

Methods

• fit(trainX: number[][], trainY: number[]): void
• predict(testX: number[][]): number[]
• predictProba(testX: number[][]): number[][]

Usage notes

• Your features represent yes/no states such as token presence, clicks, or product flags.
• You need probabilistic outputs for ranking, threshold tuning, or alert prioritization.
• You want a lightweight baseline before trying heavier linear or tree models.
• If your inputs are not already binary, decide whether to binarize upstream or use the built-in binarize threshold.
• Keep category meaning stable between training and prediction so the model sees the same feature semantics.