naivebayes

High performance implementation of the Naive Bayes algorithm in R


output: github_document

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  fig.path = "man/figures/"
)

Extended documentation can be found on the website: https://majkamichal.github.io/naivebayes/

Naïve Bayes

![BuyMeACoffee](https://img.buymeacoffee.com/button-api/?text=Buy me a coffee&emoji=&slug=michalmajka&button_colour=5F7FFF&font_colour=ffffff&font_family=Cookie&outline_colour=000000&coffee_colour=FFDD00)

CRAN_Status_Badge

Anaconda Cloud
Anaconda Cloud

1. Overview

The naivebayes package presents an efficient implementation of the widely-used Naïve Bayes classifier. It upholds three core principles: efficiency, user-friendliness, and reliance solely on Base R. By adhering to the latter principle, the package ensures stability and reliability without introducing external dependencies[1]. This design choice maintains efficiency by leveraging the optimized routines inherent in Base R, many of which are programmed in high-performance languages like C/C++ or FORTRAN. By following these principles, the naivebayes package provides a reliable and efficient tool for Naïve Bayes classification tasks, ensuring that users can perform their analyses effectively and with ease.

The naive_bayes() function is designed to determine the class of each feature in a dataset, and depending on user specifications, it can assume various distributions for each feature. It currently supports the following class conditional distributions:

  • categorical distribution for discrete features (with Bernoulli distribution as a special case for binary outcomes)
  • Poisson distribution for non-negative integer features
  • Gaussian distribution for continuous features
  • non-parametrically estimated densities via Kernel Density Estimation for continuous features

In addition to that specialized functions are available which implement:

  • Bernoulli Naive Bayes via bernoulli_naive_bayes()
  • Multinomial Naive Bayes via multinomial_naive_bayes()
  • Poisson Naive Bayes via poisson_naive_bayes()
  • Gaussian Naive Bayes via gaussian_naive_bayes()
  • Non-Parametric Naive Bayes via nonparametric_naive_bayes()

These specialized functions are carefully optimized for efficiency, utilizing linear algebra operations to excel when handling dense matrices. Additionally, they can also exploit sparsity of matrices for enhanced performance and work in presence of missing data. The package also includes various helper functions to improve user experience. Moreover, users can access the general naive_bayes() function through the excellent Caret package, providing additional versatility.

2. Installation

The naivebayes package can be installed from the CRAN repository by simply executing in the console the following line:

install.packages("naivebayes")

# Or the the development version from GitHub:
devtools::install_github("majkamichal/naivebayes")

3. Usage

The naivebayes package provides a user friendly implementation of the Naïve Bayes algorithm via formula interlace and classical combination of the matrix/data.frame containing the features and a vector with the class labels. All functions can recognize missing values, give an informative warning and more importantly - they know how to handle them. In following the basic usage of the main function naive_bayes() is demonstrated. Examples with the specialized Naive Bayes classifiers can be found in the extended documentation: https://majkamichal.github.io/naivebayes/ in this article.

3.1 Example data

library(naivebayes)

# Simulate example data
n <- 100
set.seed(1)
data <- data.frame(class = sample(c("classA", "classB"), n, TRUE),
                   bern = sample(LETTERS[1:2], n, TRUE),
                   cat  = sample(letters[1:3], n, TRUE),
                   logical = sample(c(TRUE,FALSE), n, TRUE),
                   norm = rnorm(n),
                   count = rpois(n, lambda = c(5,15)))
train <- data[1:95, ]
test <- data[96:100, -1]

3.2 Formula interface

nb <- naive_bayes(class ~ ., train)
summary(nb)

# Classification
predict(nb, test, type = "class")
nb %class% test

# Posterior probabilities
predict(nb, test, type = "prob")
nb %prob% test

# Helper functions
tables(nb, 1)
get_cond_dist(nb)

# Note: all "numeric" (integer, double) variables are modelled
#       with Gaussian distribution by default.

3.3 Matrix/data.frame and class vector

X <- train[-1]
class <- train$class
nb2 <- naive_bayes(x = X, y = class)
nb2 %prob% test

3.4 Non-parametric estimation for continuous features

Kernel density estimation can be used to estimate class conditional densities of continuous features. It has to be explicitly requested via the parameter usekernel=TRUE otherwise Gaussian distribution will be assumed. The estimation is performed with the built in R function density(). By default, Gaussian smoothing kernel and Silverman’s rule of thumb as bandwidth selector are used:

nb_kde <- naive_bayes(class ~ ., train, usekernel = TRUE)
summary(nb_kde)
get_cond_dist(nb_kde)
nb_kde %prob% test

# Class conditional densities
plot(nb_kde, "norm", arg.num = list(legend.cex = 0.9), prob = "conditional")

# Marginal densities
plot(nb_kde, "norm", arg.num = list(legend.cex = 0.9), prob = "marginal")

3.4.1 Changing kernel

In general, there are 7 different smoothing kernels available:

  • gaussian
  • epanechnikov
  • rectangular
  • triangular
  • biweight
  • cosine
  • optcosine

and they can be specified in naive_bayes() via parameter additional parameter kernel. Gaussian kernel is the default smoothing kernel. Please see density() and bw.nrd() for further details.

# Change Gaussian kernel to biweight kernel
nb_kde_biweight <- naive_bayes(class ~ ., train, usekernel = TRUE,
                               kernel = "biweight")
nb_kde_biweight %prob% test
plot(nb_kde_biweight, "norm", arg.num = list(legend.cex = 0.9), prob = "conditional")

3.4.2 Changing bandwidth selector

The density() function offers 5 different bandwidth selectors, which can be specified via bw parameter:

  • nrd0 (Silverman’s rule-of-thumb)
  • nrd (variation of the rule-of-thumb)
  • ucv (unbiased cross-validation)
  • bcv (biased cross-validation)
  • SJ (Sheather & Jones method)
nb_kde_SJ <- naive_bayes(class ~ ., train, usekernel = TRUE,
                               bw = "SJ")
nb_kde_SJ %prob% test
plot(nb_kde_SJ, "norm", arg.num = list(legend.cex = 0.9), prob = "conditional")

3.4.3 Adjusting bandwidth

The parameter adjust allows to rescale the estimated bandwidth and thus introduces more flexibility to the estimation process. For values below 1 (no rescaling; default setting) the density becomes “wigglier” and for values above 1 the density tends to be “smoother”:

nb_kde_adjust <- naive_bayes(class ~ ., train, usekernel = TRUE,
                         adjust = 0.5)
nb_kde_adjust %prob% test
plot(nb_kde_adjust, "norm", arg.num = list(legend.cex = 0.9), prob = "conditional")

3.5 Model non-negative integers with Poisson distribution

Class conditional distributions of non-negative integer predictors can be modelled with Poisson distribution. This can be achieved by setting usepoisson=TRUE in the naive_bayes() function and by making sure that the variables representing counts in the dataset are of class integer.

is.integer(train$count)
nb_pois <- naive_bayes(class ~ ., train, usepoisson = TRUE)
summary(nb_pois)
get_cond_dist(nb_pois)

nb_pois %prob% test

# Class conditional distributions
plot(nb_pois, "count", prob = "conditional")

# Marginal distributions
plot(nb_pois, "count", prob = "marginal")

  1. Specialized Naïve Bayes functions within the package may optionally utilize sparse matrices if the Matrix package is installed. However, the Matrix package is not a dependency, and users are not required to install or use it. ↩︎