Simple spam classifier using Support Vector Machine

This is a python implementation of spam classification taught in famous Machine Learning course by Andrew Ng

Train and test set are provided in spamTrain.mat, spamTest.mat

Full dataset at SpamAssassin Public Corpus

Email preprocessing pipeline

• Lower case email
• Remove HTML tags
• Replace URLs with 'httpaddr', emails with 'emailaddr', numbers with 'number', $ with 'dollar'
• Remove non-alphanumeric characters
• Tokenize
• Poster stem

Email features

Each email is represented by a N-vector. N is the number of word in dictionary (provided in vocab.txt). The dictionary is built by collecting most frequency words appear in dataset. Each ith-row in feature vector represent whether or not the ith word in dictionary appear in preprocessed email.

Training

Email features are fed to a linear SVM classifier with C parameter set to 0.1. The classifier then have to classify between two class: 0-non spam, 1-spam

Result

Train accuracy: 0.998250

Test accuracy: 0.989000

Top 15 predictors for spam classifirer

our click remov guarante visit basenumb dollar will price pleas most nbsp lo ga hour