Bernoulli Naive Bayes

Bernoulli Naive Bayes is a subcategory of the Naive Bayes Algorithm. It is typically used when the data is binary and it models the occurrence of features using Bernoulli distribution. It is used for the classification of binary features such as 'Yes' or 'No', '1' or '0', 'True' or 'False' etc. Here it is to be noted that the features are independent of one another.

Mathematics

In Bernoulli Naive Bayes model we assume that each feature is conditionally independent given the class . This means that we can calculate the likelihood of each feature occurring as:

• Here, p( |y) is the conditional probability of xi occurring provided y has occurred.
• i is the feature index
• holds binary value either 0 or 1

Now we will learn Bernoulli distribution as Bernoulli Naive Bayes works on that.

Bernoulli distribution

Bernoulli distribution is used for discrete probability calculation. It either calculates success or failure. Here the random variable is either 1 or 0 whose chance of occurring is either denoted by p or (1-p) respectively. The mathematical formula is given

Now in the above function if we put x=1 then the value of f(x) is p and if we put x=0 then the value of f(x) is 1-p. Here p denotes the success of an event.

Example:

To understand how Bernoulli Naive Bayes works, here's a simple binary classification problem.

1. Vocabulary

Extract all unique words from the training data:

Vocabulary size

2. Binary Feature Matrix (Presence = 1, Absence = 0)

Each message is represented using binary features indicating the presence (1) or absence (0) of a word.

3. Apply Laplace Smoothing

where for both classes (2 documents per class), so the denominator becomes 4.

4. Word Probabilities

For Spam class:

For Not Spam class:

5. Classify Message "buy now"

The message contains words "buy" and "now, so the feature vector is:

5.1 For Spam:

5.2 For Not Spam:

6. Final Classification

Since , the message is classified as:

Implementing Bernoulli Naive Bayes

For performing classification using Bernoulli Naive Bayes we have considered an email dataset.

The email dataset comprises of four columns named Unnamed: 0, label, label_num and text. The category of label is either ham or spam. For ham the number assigned is 0 and for spam 1 is assigned. Text comprises the body of the mail. The length of the dataset is 5171.

The dataset can be downloaded from here.

1. Importing Libraries

In the code we have imported necessary libraries like pandas, numpy and sklearn. Bernoulli Naive Bayes is a part of sklearn package.

2. Data Analysis

In this code we have performed a quick data analysis that includes reading the data, dropping unnecessary columns, printing shape of data, information about dataset etc.

Output:

(5171, 4) Index(['Unnamed: 0', 'label', 'text', 'label_num'], dtype='object')

3. Count Vectorizer

In the code since text data is used to train our classifier we convert the text into a matrix comprising numbers using Count Vectorizer so that the model can perform well.

4. Data Splitting, Model Training and Prediction

Output:

The classification report shows that for class 0 (not spam) precision, recall and F1 score are 0.84, 0.98 and 0.91 respectively. For class 1 (spam) they are 0.92, 0.56 and 0.70. The recall for class 1 drops due to the 13% spam data. The overall accuracy of the model is 86%, which is good.

Bernoulli Naive Bayes is used for spam detection, text classification, Sentiment Analysis and used to determine whether a certain word is present in a document or not.

Difference Between Different Naive Bayes Model

Here is the quick comparison between types of Naive Bayes that are Gaussian Naive Bayes, Multinomial Naive Bayes and Bernoulli Naive Bayes.