End to End NLP Pipeline for Spam Emails Classification
Spam emails classification accounts for a a great deal of tasks in internet security. This raises the need for a perfect classifier to protect agents from malware attacks.
In this project we will be processing the email raw text and get insights about the key words that most spam emails share. For the sake of maximum security, we will use two frameworks to solve the problem, TensorFlow
and NLTK
, and average to combine the results.
In this section I have proceesed the data using the NLTK framework. First, I tokenized the messages using RegexpTokenizer
. Then, I lemmatized them using WordNetLemmatizer
. Lastly, I removed the stop words and produced the tokens in its useful form.
To select the most important tokens, I generated a token counter to count the frequencies of each token in the training dataset. Then, I kept only the tokens that was repeated more than 1000 times. Now that I have the features set, I defined a counts vector to store the frequencies of the features tokens in each message. Lastly, I trained a RandomForestClassifier
and achieved 98% accuracy on the test set.
This section is found in the nltk_processing notebook.
The work is pretty much easier here. First, I fitted the Keras.preprocessing.text.Tokenizer
on the training dataset and then padded the sequences using the keras.preprocessing.sequence.pad_sequences
to produce inputs with equal lengths.
I have constructed a simple DNN and fitted it on the data. The model achieved 99% accuracy on the validation datset.
This section is found in the tensorflow_processing notebook.
Lastly, we used a simple webpage to deploy the models. We used the flask
API to render the HTML webpage and provided the the index file with the data field result storing the results from the models and the probability for each model from the predict_spam
function in the app file. For the sake of better visualization, we used a CSS style file to produce the webpage in its final version as seen above.