Build your own Spam Classifier using Naive Bayes - Machine Learning.

In the initial phase of machine learning, you're more likely to deal with classification models. Sometimes maybe you'll wonder which project is the best to start the journey toward classification problems. Spam classification and Digit recognition are two of the projects both beginners and experts consider since it gives you a basic idea of how things are done in classification.

Let's create a Spam classifier using Naive Bayes and Tf-IDF Vectorizer.

Loading the Dataset

Dataset can be downloaded from Kaggle. Lots of datasets are available in Kaggle of different messages and the type you may receive sometimes.

#Loading the dataset
import pandas as pd
df = pd.read_csv('sms_spam.csv')
df

df['type'].value_counts()

ham     4812
spam     747
Name: type, dtype: int64

Dataset can be fetched by the read_csv function in pandas. This particular dataset contains 5559 rows of 747 Spam and 4812 Ham messages. However,  one thing is missing, we need the labels in numbers, for instance, we can add another column to the dataset containing 0 and 1 for ham and spam respectively.

# assigning labels to the dataset, 0 for ham and 1 for spam
df['label_num'] = df['type'].apply(lambda x: 1 if x == 'spam' else 0)

Now if you run this code it will create a new column at the end with the specific labels for each message.

Splitting the dataset

Usually, we split the entire dataset into train and test data for the model to perform. The model is trained by the training data and its corresponding labels. Then it can be tested with the test data to check whether the model is good at predicting or not.

#splitting the dataset
from sklearn.model_selection import train_test_split
X_train, X_test, y_train, y_test = train_test_split(df['text'], df['label_num'], test_size=0.3, random_state=42)

 rows in train and test set -

print("train set:", X_train.shape)  # rows in train set
print("test set:", X_test.shape)  # rows in test set

Tf-IDF Vectorization

Tf-IDF is one of the efficient statistical methods to figure out the words that are relevant in a text, sentence, or paragraph. We know that the messages are texts, however, computers aren't good at texts so Tf-IDF helps to convert the texts to corresponding numerical values so as to fit the data to the model.

from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.naive_bayes import MultinomialNB
# Converting X_train to a list
lst = X_train.tolist()
# Applying Tf-IDF Vectorization
vectorizer = TfidfVectorizer(input = lst, lowercase = True, stop_words = "english")
train_transformed = vectorizer.fit_transform(X_train)
test_transformed = vectorizer.transform(X_test)
# Fit the transformed train data to the model.
model = MultinomialNB()
model.fit(train_transformed, y_train)

Here we applied the Tf-IDF vectorizer and fitted the transformed train data to the Multinomial Naive Bayes classifier. Now let's look at the predictions made by the model.

Run this code to see the predictions and compare them with actual values. 

prediction = model.predict(test_transformed)
actual = y_test
print("Prediction:", list(prediction))
print("Actual:    ",list(actual))

Evaluating the Model

Confusion matrix:-

from sklearn.metrics import confusion_matrix
matrix = confusion_matrix(prediction, actual)
matrix

array([[1457,   52],
       [   0,  159]], dtype=int64)

precision = matrix[1][1]/(matrix[1][1]+matrix[0][1])
recall = matrix[1][1]/(matrix[1][1]+matrix[1][0])
f1score = matrix[1][1]/(matrix[1][1]+(matrix[1][0]+(matrix[0][1]/2)))
print("precision score:", precision)
print("recall score:", recall)
print("f1_score:", f1score)

precision score: 0.7535545023696683
recall score: 1.0
f1_score: 0.8594594594594595

The scores are not bad at all. Now if you look at the recall score(sensitivity) and the precision score is 1.0 and 0.75.. respectively. The recall is so high that the model can identify most of the True Positives. This means the model is able to classify most of the spam messages including some ham messages. However, the precision and f1 scores are also not bad. 

Let's predict some real messages. Here are some messages that I received in the past.

Congragulations! You have won a $10,000. Go to https://bit.ly/23343 to claim now.
Get $10 Amazon Gift Voucher on Completing the Demo:- va.pcb3.in/ click this link to claim now
You have won a $500. Please register your account today itself to claim this offer https://imp.com.
Please dont respond to missed calls from unknown international numbers Call/ SMS on winning prize. lottery as this may be fraudulent call.

messages = ["Congragulations! You have won a $10,000. Go to https://bit.ly/23343 to claim now.",
           "Get $10 Amazon Gift Voucher on Completing the Demo:- va.pcb3.in/ click this link to claim now",
           "You have won a $500. Please register your account today itself to claim now https://imp.com",
           "Please dont respond to missed calls from unknown international numbers Call/ SMS on winning prize. lottery as this may be fraudulent call."
          ]
message_transformed = vectorizer.transform(messages)
new_prediction = model.predict(message_transformed)
for i in range(len(new_prediction)):
    if new_prediction[i] == 0:
        print("Ham.")
    else:
        print("Spam.")

Spam.
Spam.
Spam.
Ham.

The first three messages I received were spam and the last one is probably ham. The model prediction is correct when I run this code. However, this model may not be that perfect for a real-life application since this is a basic understanding of spam classification. but you can optimize the model with some of the performance measures like Grid Search. Try to collect more datasets and work with some of the good datasets available there on platforms like Kaggle.