DocumentFeatureSelection 0.8rc3

Various methods of feature selection from Text Data

DocumentFeatureSelection

what’s this?

This is set of feature selection codes from text data. (About feature selection, see here or here)

The Feature selection is really important when you use machine learning metrics on natural language data. The natural language data usually contains a lot of noise information, thus machine learning metrics are weak if you don’t process any feature selection. (There is some exceptions of algorithms like Decision Tree or Random forest . They have feature selection metric inside the algorithm itself)

The feature selection is also useful when you observe your text data. With the feature selection, you can get to know which features really contribute to specific labels.

Please visit project page on github.

If you find any bugs and you report it to github issue, I’m glad.

Any pull-requests are welcomed.

Supporting methods

This package provides you some feature selection metrics. Currently, this package supports following feature selection methods

• TF-IDF

• Pointwise mutual information (PMI)

• Strength of Association (SOA)

• Bi-Normal Separation (BNS)

Contribution of this package

• Easy interface for pre-processing

• Easy interface for accessing feature selection methods

• Fast speed computation thanks to sparse matrix and multi-processing

Overview of methods

TF-IDF

This method, in fact, just calls TfidfTransformer of the scikit-learn.

See scikit-learn document about detailed information.

PMI

PMI is calculated by correlation between feature (i.e. token) and category (i.e. label). Concretely, it makes cross-table (or called contingency table) and calculates joint probability and marginal probability on it.

To know more, see reference

In python world, NLTK and Other package also provide PMI. Check them and choose based on your preference and usage.

SOA

SOA is improved feature-selection method from PMI. PMI is weak when feature has low word frequency. SOA is based on PMI computing, however, it is feasible on such low frequency features. Moreover, you can get anti-correlation between features and categories.

In this package, SOA formula is from following paper,

Saif Mohammad and Svetlana Kiritchenko, "Using Hashtags to Capture Fine Emotion Categories from Tweets", Computational Intelligence, 01/2014; 31(2).

SOA(w, e)\ =\ log_2\frac{freq(w, e) * freq(\neg{e})}{freq(e) * freq(w, \neg{e})}

Where

• freq(w, e) is the number of times w occurs in an unit(sentence or document) with label e

• freq(w,¬e) is the number of times w occurs in units that does not have the label e

• freq(e) is the number of units having the label e

• freq(¬e) is the number of units having NOT the label e

BNS

BNS is a feature selection method for binary class data. There is several methods available for binary class data, such as information gain (IG), chi-squared (CHI), odds ratio (Odds).

The problem is when you execute your feature selection on skewed data. These methods are weak for such skewed data, however, BNS is feasible only for skewed data. The following paper shows how BNS is feasible for skewed data.

Lei Tang and Huan Liu, "Bias Analysis in Text Classification for Highly Skewed Data", 2005

or

George Forman, "An Extensive Empirical Study of Feature Selection Metrics for Text Classification",Journal of Machine Learning Research 3 (2003) 1289-1305

Requirement

• Python 3.x(checked under Python 3.5)

Setting up

install

python setup.py install

Examples

See scripts in examples/

Change log

0.6 2016/04/02

supports PMI and TF-IDF under Python3.x

0.7 2016/04/03

Added SOA under Python3.x

0.8 2016/04/03

Added BNS under Python3.x