In this research, we study the performances of two different models for spam filtering. Specifically, we use the scikit-learn pre-implemented models for Naive Bayes and Support Vector Machine (SVM) and evaluate the results on the TReC 2007 Spam Dataset.

To create an environment with the environment.yml file, navigate to the folder containing the environment file and type conda env create -f environment.yml into your terminal. To activate your environment, do conda activate env-name. The environment in the yml file is named IR.

Dependency

sys os pickle csv time

pandas numpy collections re

nltk nltk.corpus nltk.stem rake

keras keras.utils keras.callbacks

sklearn

Process Trec 2007 Spam Dataset for spam filtering model

• Decoding Email: DataProcessing.ipynb
  • Decode email of byte type to string type
• Extract Contents: DataPreProcessing.ipynb
  • Extracts sender and subject of each email
  • Filter email contents that is not English word and not English stop words by nltk corpus
  • Save to txt file which contains each filtered email as a line

• Evaluation of Models: eval_utils.py
  • This program provides function evaluate that prints Precision, Recall, F1 Score, and Accuracy of the given model
• Models for Spam Detection: Naive_Bayes.ipynb
  • This file contains three different Naive Bayes model and Support Vector Machine (SVM)
    • Gaussian Naive Bayes
    • Multinomial Naive Bayes
    • Complement Naive Bayes
  • We used 67% of dataset for training and 33% for evaluation

• Gordon V. Cormack (2007) TREC 2007 Spam Track Overview