Spam Detection using Machine Learning

What we are building

A spam detector is a machine learning program that reads a message and decides if it is spam or not. It learns from example messages already labelled as spam or safe. This is a classic supervised classification project, built easily with the scikit-learn library.

How spam detection works

Spam messages tend to use certain words more often ("free", "win", "offer"). The model looks at many labelled examples, learns which words point to spam, and uses that to judge new messages. Because it learns from labelled data, it is supervised learning — specifically, classification into two classes: spam or not spam.

The data

messages = [
    "Win a free prize now",       # spam
    "Meeting at 5 pm today",      # not spam
    "Claim your free reward",     # spam
    "Lunch tomorrow?"             # not spam
]
labels = [1, 0, 1, 0]   # 1 = spam, 0 = not spam

We need messages and their labels. Here 1 means spam and 0 means not spam. Real projects use thousands of examples, but the idea is the same.

Turning text into numbers

from sklearn.feature_extraction.text import CountVectorizer

vec = CountVectorizer()
X = vec.fit_transform(messages)   # word counts as numbers

Models cannot read words, so CountVectorizer turns each message into a row of word counts. Now every message is a list of numbers the model can learn from.

Training the model

from sklearn.naive_bayes import MultinomialNB

model = MultinomialNB()
model.fit(X, labels)   # learn from the examples

MultinomialNB (Naive Bayes) is a fast, reliable model for text. fit() trains it on our numbers and labels, so it learns which words signal spam.

Making predictions

new = ["Free reward waiting for you"]
new_X = vec.transform(new)         # same vectorizer
print(model.predict(new_X))        # [1] = spam

We transform the new message with the same vectorizer, then call predict. A result of 1 means the model thinks it is spam. That is a working spam detector.

Common mistakes

• Using a different vectorizer for new data. Use transform with the same fitted vectorizer, not a new one.
• Too little data. A few examples cannot learn well — real models use many messages.
• Not splitting train and test. Check accuracy on unseen data to avoid fooling yourself.
• Forgetting to clean text. Lowercasing and removing punctuation usually improves results.

Summary

• Spam detection is supervised classification: spam or not spam.
• Turn text into numbers with CountVectorizer (or TF-IDF).
• Train a MultinomialNB model with fit().
• Predict new messages with the same vectorizer and predict().
• Use plenty of data and a train/test split for a reliable model.

Next, understand the ML type in supervised learning, or check model quality with overfitting and underfitting.