Spam Classification Problem for LA Coding Dojo

Tasks:

1. Process data into Weka ARFF format
2. Train a classifier using Weka
3. Test your classifier against test data

DATA PROCESSING

• Download training data tarball from http://dojo.v.wc1.atti.com/data/train
• Weka accepts ARFF files: http://www.cs.waikato.ac.nz/ml/weka/arff.html
  • See https://github.com/a34729t/coding-dojo-ml/blob/master/sample.arff
• Each training or test file becomes a vector of features
• You can play around with the values or ranges of values for your features in the ARFF file itself

TRAINING

• POST to http://dojo.v.wc1.atti.com/classify/train with the following params:
  • name (your name)
  • classifier (weka classifier name, see list below)
  • featuresTrain (your ARFF file)
• See https://github.com/a34729t/coding-dojo-ml/blob/master/client.rb for a sample client
• You get back a JSON response with statistics on how well you performed
• Check your performance vs others at http://dojo.v.wc1.atti.com/rank

TESTING

• After a certain period of time we will allow you to access the training data
• Download test data tarball from http://dojo.v.wc1.atti.com/data/test
• Process the test data with the same features you use for training data
• POST to http://dojo.v.wc1.atti.com/classify/train with params:
  • name
  • classifier
  • featuresTrain
  • featuresTest

NOTES

• I would recommend using Naive Bayes and the Decision Tree first as they give you good feedback about which features are useful
• Machine Learning is about quality data and features—what model you use is much less important

WEKA CLASSIFIERS (From simplest to more complicated)

• weka.classifiers.bayes.NaiveBayes
• weka.classifiers.trees.J48
• weka.classifiers.functions.Logistic
• weka.classifiers.lazy.kstar
• weka.classifiers.rules.JRip
• weka.classifiers.functions.SMO