This project has been carried on as part of the final assignment of the L101 module. It aims at studying Memory-Based and Naive Bayes models for Spam Filtering.
It is composed of multiple elements:
- a PDF report that describes all the work done, presents results and offers analyses about models and results
- a jupyter notebook spam_filtering.ipynb that gathers all the code used to preprocessed data and run models
- a jupyter notebook results.ipynb that gathers plots to visualise and compare results
- a directory GenSpam that gathers data from the GenSpam dataset introduced by Ben Medlock
- a csv file timbl_results.csv that gathers results of all experiments with Memory-Based models
- a csv file nb_results.csv that gathers results of all experiments with Naive-Bayes models
Memory-Based models are run thanks to TiMBL Software (Daelemans et al., 2000). To do so, we run in our console the command line: timbl -f data_timbl/data_700.train -t data_timbl/data_700.test -wgr -dID -k1 +vcs
This command line is composed of the following arguments:
-f: file with the train set in the C4.5 format (see Section 2.2 in the report)-t: file with the adaptation/test set in the C4.5 format-w: feature-weighting scheme (gr: Gain Ratio, 0: Equal Weights)-d: distance-weighting scheme (z: Equal Distance, ID: Inverse Distance, IL: Inverse Linear, ED: Exponential Decay)+v: output format (cs: class statistics with precision, recall, f1-score and AUC metrics)
Models and parameters selection are carried out with the train and adaptation sets, while methods comparison is realised with the train and test sets.
Naive Bayes models are run through to their scikit-learn implementation for Python. Corresponding code can be found in the spam_filtering.ipynb notebook.