The 20 Newsgroups data set is a collection of approximately 20,000 newsgroup documents, partitioned (nearly) evenly across 20 different newsgroups. Download link: http://qwone.com/~jason/20Newsgroups/

Selected 4 classes for this project.

1. rec.sport.hockey
2. sci.med
3. soc.religion.christian
4. talk.religion.misc

Classifiers used - Naive Bayes, Logistic Regression, Support Vector Machines, and Random Forests.

Configurations used:

1. Unigram Baseline (UB) -- Basic sentence segmentation and tokenization. Use all words.
2. Bigram Baseline (BB) -- Use all bigrams. (e.g. I ran a race => {I ran, ran a, a race}. ) Applied all the classifiers for these configurations and selected best model from that and applied some more techniques namely:

1. Feature representations
2. Feature selection
3. Hyperparameters

My Best configuration is obtained by removing stop words with L2 penalization on SVM. I have used SGDclassifier with hinge loss ( a linear SVM, as the number of features is more than 10000 most likely points are linearly separable ) as the classifier with above mentioned. I have created two python files one to build a model and another to test the model:

Execute the following command to get a learned model from a training dataset:

python best_config_train.py Testing_Samples_Location.

Execute the following command to pre-learned model run on a test dataset:

python best_config_test.py Testing_Samples_Location

Exploration for best params in Unigram Configuration:

python Analysis_config.py Training_samples_location Testing_Samples_Location

Code for Learning curves plot: python Learning_curves.py

Baseline models: python Unigram_bigram_models.py Training_samples_location Testing_Samples_Location