Skip to main content

Spam Detection


Nowadays, many e-mail are being sent and many of them are spam e-mails. In this tutorial, I will guide you through the steps of building a small spam detection application (written in Python).

Data scientist like this Python library: Scikit-learn. This library is capable of doing all kinds of Machine Learning magic. So, make sure that you install this library first.


The spam detection problem is in fact a text classification problem. An e-mail (a text document) is either “spam” or “no spam”. In text mining, this is called single-label text classification, since there is only one label: “spam”. A classifier is an algorithm that is capable of telling whether a text document is either “spam” or “no spam”. In this article, we first setup a small dataset on which we will train a classifier. Then, we will create a test dataset and view the results. Before we can do all of this, we need to extract features from the texts.

Feature extraction

Take a look at the following e-mails:

$1,000 ALARMING!

An appointment on July 2nd

New course content for Text Classification


It is hopefully clear that the first and the last e-mail are definitely spam. The third e-mail can be viewed as spam, but since we can make decisions, we make the decision that the second and the third e-mail are not spam.

First of all, do the following imports in Python:

import numpy as np
from sklearn.linear_model import SGDClassifier

So, the burning question is which features should we use? It seems that e-mails with many “!”, “$” or capitals are probably spam e-mails. So, lets create 3 features and lets create our training and test set. The feature vector consists of three entries where the first entry is 1 if there is a “!” in the text (a 0 otherwise). The second entry is 1 if there is a “$” in the text (0 otherwise) and the last entry is the ratio of uppercase characters with respect to the sum of the number of uppercase characters and the number of lowercase characters.

In Python we can create a training set and a test set as follows (the “isspam” variables define whether the text is spam or not):

# Define a training set
training_data = ["$1,000 ALARMING!", "An appointment on July 2nd", "New course content for Text Classification", "BUY VIAGRA!!"]
training_isspam = [True, False, False, True]

# Define the testing set
testing_data = ["New course content for Information Retrieval", "MAkE $$$!", "Grades available for Text Mining", "SELL HOUSE FOR $1,000,000!!"]
testing_isspam = [False, True, False, True]

For the feature extraction, we can write the following method:

def extract_features(text):
    Extract features from a given text.

    :param text: Text to extract features for.
    :return:     A vector where:
                    - The 0th element is 1 if there is a "!" inside the text (0 otherwise).
                    - The 1th element is 1 if there is a "$" inside the text (0 otherwise).
                    - The 2nd element is the ratio of uppercase characters with respect to the sum of all uppercase and
                      lowercase characters.
    features = np.zeros((3,))
    if "!" in text:
        features[0] = 1
    if "$" in text:
        features[1] = 1
    # A list consisting of lowercase characters
    lowercase = list('abcdefghijklmnopqrstuvwxyz')
    # A list consisting of uppercase characters
    # Set the counts of lowercase and uppercase characters to 0
    num_lowercase = 0
    num_uppercase = 0
    # And count the lowercase and uppercase characters
    for char in text:
        if char in lowercase:
            num_lowercase += 1
        elif char in uppercase:
            num_uppercase += 1
    # Define the third feature as the ratio of uppercase characters
    features[2] = num_uppercase / (num_lowercase + num_uppercase)
    return features

Text Classification

What to do next now we have translated the text into numerical values? First note that in our case, our numerical values per document are in a three dimensional space: there are three dimensions/features. A linear classifier tries to separate datapoints in this space by placing a so called hyperplane. You can think of an hyperplane as a line or a plane in higher dimensions. Everything below the hyperplane is classified as spam and everything above the hyperplane is classified as not-spam (or the otherway around, just how you define the hyperplane). Now, there are many choices for a Text Classification algorithm. A choice that does work well in simple cases is the Stochastic Gradient Descent classifier. In Python, it is easy to train the classifier on the training set and test it on the test set. This can be done as follows:

# Make an array of features where the ith row corresponds to the ith documents and the columns correspond to the features
training_features = np.vstack([extract_features(training_data[i]) for i in range(len(training_data))])

# Make an Stochastic Gradient Descent classifier
clf = SGDClassifier()
# And fit it to the training set, training_isspam)

# Predict the labels of the test set
for test_index in range(len(testing_data)):
    features = extract_features(testing_data[test_index])
    print("Test case:")
    print(20 * '=')
    print('Text:', 4 * "t", testing_data[test_index])
    print('Features:', 3 * "t", features)
    print('Predicted is spam:', "t", clf.predict(features))
    print('Is spam:', 3 * "t", testing_isspam[test_index])


After executing the code, we get the following results:

Test case:
Text: 				 New course content for Information Retrieval
Features: 			 [ 0.          0.          0.07692308]
Predicted is spam: 	         False
Is spam: 			 False

Test case:
Text: 				 MAkE $$$!
Features: 			 [ 1.    1.    0.75]
Predicted is spam: 	         True
Is spam: 			 True

Test case:
Text: 				 Grades available for Text Mining
Features: 			 [ 0.          0.          0.10714286]
Predicted is spam: 	         False
Is spam: 			 False

Test case:
Text: 				 SELL HOUSE FOR $1,000,000!!
Features: 			 [ 1.  1.  1.]
Predicted is spam: 	         True
Is spam: 			 True

As you can see, the classifier does the right thing! Now we can classify little spammy e-mail messages.


Try to implement more features, like the words used in text messages. For example, the word “VIAGRA” is often used in spam e-mails. Try to classify some more text messages. If you need any help, you can send me a (not so spammy) e-mail message.

Kevin Jacobs

Kevin Jacobs

Kevin Jacobs is a certified Data Scientist and blog writer for Data Blogger. He is passionate about any project that involves large amounts of data and statistical data analysis. Kevin can be reached using Twitter (@kmjjacobs), LinkedIn or via e-mail: