yassersouri/classify-text

Salam

Text Classification with python

This is an experiment. We want to classify text with python.

Dataset

For dataset I used the famous "Twenty Newsgrousps" dataset. You can find the dataset freely here.

I've included a subset of the dataset in the repo, located at dataset\ directory. This subset includes 6 of the 20 newsgroups: space, electronics, crypt, hockey, motorcycles and forsale.

When you run main.py it asks you for the root of the dataset. You can supply your own dataset assuming it has a similar directory structure.

UTF-8 incompatibility

Some of the supplied text files had incompatibility with utf-8!

Even textedit.app can't open those files. And they created problem in the code. So I'll delete them as part of the preprocessing.

Requirements

python 2.7
python modules:
- scikit-learn (v 0.11)
- scipy (v 0.10.1)
- colorama
- termcolor
- matplotlib (for use in plot.py)

The code

The code is pretty straight forward and well documented.

Running the code

python main.py

Experiments

For experiments I used the subset of the dataset (as described above). I assume that we like hockey, crypt and electronics newsgroups, and we dislike the others.

For each experiment we use a "feature vector", a "classifier" and a train-test splitting strategy.

Experiment 1: BOW - NB - 20% test

In this experiment we use a Bag Of Words (BOW) representation of each document. And also a Naive Bayes (NB) classifier.

We split the data, so that 20% of them remain for testing.

Results:

             precision    recall  f1-score   support

   dislikes       0.95      0.99      0.97       575
      likes       0.99      0.95      0.97       621

avg / total       0.97      0.97      0.97      1196

Experiment 2: TF - NB - 20% test

In this experiment we use a Term Frequency (TF) representation of each document. And also a Naive Bayes (NB) classifier.

We split the data, so that 20% of them remain for testing.

Results:

             precision    recall  f1-score   support

   dislikes       0.97      0.92      0.94       633
      likes       0.91      0.97      0.94       563

avg / total       0.94      0.94      0.94      1196

Experiment 3: TFIDF - NB - 20% test

In this experiment we use a TFIDF representation of each document. And also a Naive Bayes (NB) classifier.

We split the data, so that 20% of them remain for testing.

Results:

             precision    recall  f1-score   support

   dislikes       0.96      0.95      0.95       584
      likes       0.95      0.96      0.96       612

avg / total       0.95      0.95      0.95      1196

Experiment 4: TFIDF - SVM - 20% test

In this experiment we use a TFIDF representation of each document. And also a linear Support Vector Machine (SVM) classifier.

We split the data, so that 20% of them remain for testing.

Results:

             precision    recall  f1-score   support

   dislikes       0.96      0.97      0.97       587
      likes       0.97      0.96      0.97       609

avg / total       0.97      0.97      0.97      1196

Experiment 5: TFIDF - SVM - KFOLD

In this experiment we use a TFIDF representation of each document. And also a linear Support Vector Machine (SVM) classifier.

We split the data using Stratified K-Fold algorithm with k = 5.

Results:

Mean accuracy: 0.977 (+/- 0.002 std)

Experiment 5: BOW - NB - KFOLD

In this experiment we use a TFIDF representation of each document. And also a linear Support Vector Machine (SVM) classifier.

We split the data using Stratified K-Fold algorithm with k = 5.

Results:

Mean accuracy: 0.968 (+/- 0.002 std)

Experiment 6: TFIDF - SVM - 90% test

In this experiment we use a TFIDF representation of each document. And also a linear Support Vector Machine (SVM) classifier.

We split the data, so that 90% of them remain for testing! Only 10% of the dataset is used for training!

Results:

             precision    recall  f1-score   support

   dislikes       0.90      0.95      0.93      2689
      likes       0.95      0.90      0.92      2693

avg / total       0.92      0.92      0.92      5382

Experiment 7: TFIDF - SVM - KFOLD - 20 classes

In this experiment we use a TFIDF representation of each document. And also a linear Support Vector Machine (SVM) classifier.

We split the data using Stratified K-Fold algorithm with k = 5.

We also use the whole "Twenty Newsgroups" dataset, which has 20 classes.

Results:

Mean accuracy: 0.892 (+/- 0.001 std)

Experiment 7: BOW - NB - KFOLD - 20 classes

In this experiment we use a Bag Of Words (BOW) representation of each document. And also a Naive Bayes (NB) classifier.

We split the data using Stratified K-Fold algorithm with k = 5.

We also use the whole "Twenty Newsgroups" dataset, which has 20 classes.

Results:

Mean accuracy: 0.839 (+/- 0.003 std)

Experiment 8: TFIDF - 5-NN - Distance Weights - 20% test

In this experiment we use a TFIDF representation of each document. And also a K Nearest Neighbors (KNN) classifier with k = 5 and distance weights.

We split the data using Stratified K-Fold algorithm with k = 5.

Results:

             precision    recall  f1-score   support

   dislikes       0.93      0.88      0.90       608
      likes       0.88      0.93      0.90       588

avg / total       0.90      0.90      0.90      1196

Experiment 9: TFIDF - 5-NN - Uniform Weights - 20% test

In this experiment we use a TFIDF representation of each document. And also a K Nearest Neighbors (KNN) classifier with k = 5 and uniform weights.

We split the data using Stratified K-Fold algorithm with k = 5.

Results:

             precision    recall  f1-score   support

   dislikes       0.95      0.90      0.92       581
      likes       0.91      0.95      0.93       615

avg / total       0.93      0.93      0.93      1196

Experiment 10: TFIDF - 5-NN - Distance Weights - KFOLD

In this experiment we use a TFIDF representation of each document. And also a K Nearest Neighbors (KNN) classifier with k = 5 and distance weights.

We split the data using Stratified K-Fold algorithm with k = 5.

Results:

Mean accuracy: 0.908 (+/- 0.003 std)

Experiment 11: TFIDF - 5-NN - Distance Weights - KFOLD - 20 classes

In this experiment we use a TFIDF representation of each document. And also a K Nearest Neighbors (KNN) classifier with k = 5 and distance weights.

We split the data using Stratified K-Fold algorithm with k = 5.

We also use the whole "Twenty Newsgroups" dataset, which has 20 classes.

Results:

 Mean accuracy: 0.745 (+/- 0.002 std)

So What?

This experiments show that text classification can be effectively done by simple tools like TFIDF and SVM.

Any Conclusion?

We have found that TFIDF with SVM have the best performance.

TFIDF with SVM perform well both for 2-class problem and 20-class problem.

I would say if you want suggestion from me, use TFIDF with SVM.