wnb 0.2.6

Python library for the implementations of general and weighted naive Bayes (WNB) classifiers.

General and weighted naive Bayes classifiers

Scikit-learn-compatible

Introduction

Naive Bayes is often recognized as one of the most popular classification algorithms in the machine learning community. This package takes naive Bayes to a higher level by providing its implementations in more general and weighted settings.

General naive Bayes

The issue with the well-known implementations of the naive Bayes algorithm (such as the ones in sklearn.naive_bayes module) is that they assume a single distribution for the likelihoods of all features. Such an implementation can limit those who need to develop naive Bayes models with different distributions for feature likelihood. And enters WNB library! It allows you to customize your naive Bayes model by specifying the likelihood distribution of each feature separately. You can choose from a range of continuous and discrete probability distributions to design your classifier.

Weighted naive Bayes

Although naive Bayes has many advantages such as simplicity and interpretability, its conditional independence assumption rarely holds true in real-world applications. In order to alleviate its conditional independence assumption, many attribute weighting naive Bayes (WNB) approaches have been proposed. Most of the proposed methods involve computationally demanding optimization problems that do not allow for controlling the model's bias due to class imbalance. Minimum Log-likelihood Difference WNB (MLD-WNB) is a novel weighting approach that optimizes the weights according to the Bayes optimal decision rule and includes hyperparameters for controlling the model's bias. WNB library provides an efficient implementation of gaussian MLD-WNB.

Installation

The easiest way to install the wnb library is by using pip:

pip install wnb

This library is shipped as an all-in-one module implementation with minimalistic dependencies and requirements. Furthermore, it fully adheres to Scikit-learn API ❤️.

Getting started ⚡️

Here, we show how you can use the library to train general and weighted naive Bayes classifiers.

General naive Bayes

A general naive Bayes model can be set up and used in four simple steps:

1. Import the GeneralNB class as well as Distribution enum class

from wnb import GeneralNB, Distribution as D

2. Initialize a classifier and specify the likelihood distributions

gnb = GeneralNB(distributions=[D.NORMAL, D.CATEGORICAL, D.EXPONENTIAL])

3. Fit the classifier to a training set (with three features)

gnb.fit(X, y)

4. Predict on test data

gnb.predict(X_test)

Weighted naive Bayes

An MLD-WNB model can be set up and used in four simple steps:

1. Import the GaussianWNB class

from wnb import GaussianWNB

2. Initialize a classifier

wnb = GaussianWNB(max_iter=25, step_size=1e-2, penalty="l2")

3. Fit the classifier to a training set

wnb.fit(X, y)

4. Predict on test data

wnb.predict(x_test)

Compatibility with Scikit-learn 🤝

The wnb library fully adheres to the Scikit-learn API, ensuring seamless integration with other Scikit-learn components and workflows. This means that users familiar with Scikit-learn will find the WNB classifiers intuitive to use.

Both Scikit-learn classifiers and WNB classifiers share these well-known methods:

• fit(X, y)
• predict(X)
• predict_proba(X)
• predict_log_proba(X)
• score(X, y)
• get_params()
• set_params(**params)
• etc.

By maintaining this consistency, WNB classifiers can be easily incorporated into existing machine learning pipelines and processes.

Benchmarks 📊

We conducted benchmarks on three datasets, Breast Cancer, Digits, and Wine, to evaluate the performance of WNB classifiers and compare them with their Scikit-learn counterpart, GaussianNB. The results show that WNB classifiers generally perform better in certain cases.

Dataset	Scikit-learn Classifier	Accuracy	WNB Classifier	Accuracy
Breast Cancer	GaussianNB	0.939	GaussianWNB	0.951
Digits	GaussianNB	0.838	GeneralNB	0.889
Wine	GaussianNB	0.974	GeneralNB	0.981

These benchmarks highlight the potential of WNB classifiers to provide better performance in certain scenarios by allowing more flexibility in the choice of distributions and incorporating weighting strategies.

The benchmark scripts used to obtain these results can be found under tests/benchmarks/ directory.

Tests

To run the tests, make sure to clone the repository and install the development requirements:

pip install -r requirements_dev.txt

Then, run pytest:

pytest

Support us 💡

You can support the project in the following ways:

⭐ Star WNB on GitHub (click the star button in the top right corner)

💡 Provide your feedback or propose ideas in the Issues section

📰 Post about WNB on LinkedIn or other platforms

Citation 📚

If you utilize this repository, please consider citing it with:

@misc{wnb,
  author = {Mohammd Mehdi Samsami},
  title = {WNB: General and weighted naive Bayes classifiers},
  year = {2023},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/msamsami/weighted-naive-bayes}},
}