AntiSpam NLP Pipeline

Index of contents

1. Objective
2. About the data
3. Research Theory behind Decision-Making Process 🚧
   • EDA
   • Preprocessing
   • Experimentation Theory behind Decision-Making Process 🚧
4. Technologies 🚧
5. Installation 🚧

Project Development Roadmap

This project started long time ago as a data science project for my personal portfolio with the goal of gaining experience in working with NLP techniques for text classification but I decided to make it grow and create a full implemented project with Machine Learning, Deep Learning (LLMs) and MLOps methods.

About the data

This project is based on four datasets extracted from Kaggle

• Spam Email with 5,572 rows
• Spam Email Classification Dataset with 83446 rows
• Email Classification (Ham-Spam) with 179 rows
• Spam email Dataset with 5695 rows

All datasets have a similar format consisting on two columns:

1. Message/text/email that contains the emails
2. Message/label/spam that contains ham/spam or 0/1 values

Due to lfs constrains, none of the datasets nor the concatenated raw dataset created in the notebook inside 01_feature_pipeline

Research

EDA

Notebook

1. Basic overview of the training set features.
2. Analysis of different approaches for feature engineering.
3. Analysis of different approaches for data cleaning.
4. Basic viz of features.

Preprocessing

Notebook

1. Change of data types

2. Drop duplicates

3. Chars cleaning. Based on this Notebook

   1. Replacement of special replacements
   2. Replacement of emojis
   3. Conversion to lowecase
   4. Removal of HTML tags
   5. Removal of URLs
   6. Replacement of numbers with "number"
   7. Replacement of e-mail addresses with "emailaddr"
   8. Removal of punctutation
   9. Removal of Non-Alphabetic Characters
   10. Collapse of multiple whitespaces into single whitespace

4. Tokenization

5. Removal of stopwords

6. Lemmatization

Experimentation

This section only contains results. If you want to take a look at the Decision-Making Process please take a look at the Experimentation README file

1. Base model BOW + Multinomial Naive Bayes

Classification Report (Test):

              precision    recall  f1-score   support

       0       0.99      0.97      0.98       453
       1       0.82      0.95      0.88        63

accuracy                           0.97       516
macro avg      0.91      0.96      0.93       516
weighted avg   0.97      0.97      0.97       516

• Recall 0.95: the model is detecting a 95% of the class 1

• Precision 0.82: lower than we would like to because there is a significative number of False Positives

• F1-score 0.88: decent value but since recall is high and precission not that much there is room for improvement.

Technologies

• Docker

• aws

Installation

build docker container

docker build -t mlflow .

run docker container

docker run -p 5000:5000 -v $(pwd)/mlflow_artifacts:/mlflow/artifacts mlflow

Conclusions