S-RANK

A basic implementation of the Scalable RANK Algorithm, for feature selection in unsupervised learning problems, as described in this article by Manoranjan Dash and Huan Liu.

Description

All the theoretical details are presented inside the article above. I've implemented the RANK and SRANK algorithm following its indications.

Dependencies

pandas v1.0.1
scikit-learn v0.2

Usage

In order to use the algorithm, install the module with:

pip install S-RANK

Then, to import:

from S-RANK.FeatureRanker import SRANK

Then, in order to use the algorithm:

ranker = SRANK()
ranker.apply(df_big, vars_type, discrete_var_list = None, clean_bool = False, rescale_bool = False, 
             N_SAMPLE = 40, SAMPLE_SIZE = 50)

Arguments:

df_big: The entire dataframe to be analyzed. It has to be a pandas.Dataframe object. Please make sure that all variable dtypes are inferred correctly. It is important to have numeric variables and non-numeric variables (dtype: object) inferred correctly. Up to now, the algorithm is able to handle numeric and object data, any other type (like datetimes) will not be treated. It is recommended to remove any non-numerical and non-object feature.
vars_type : the type of data in the dataframe. String that can take 3 values: continous if all the variables have continous numerical values. discrete if all the variables are categorical (encoded as string or number makes no difference), mixed if there's both.
discrete_vars_list : needs to be set only if vars_type = "mixed" . It is a list containing the names of the categorical variable in the dataframe. If vars_type = "continous" or discrete this must be set to None.
clean_bool if this option is set to true, outliers removal is performed. Any instance (row) is identified as an outlier if any of its feature lays outside the interval M +- 3 * s where M is the mean value of the feature and s is its standard deviation.
rescale_bool if this option is set to True, rescaling of continous variables is performed, i.e., all values are rescaled to be inside the interval [0;1]. This affects only continous variables, since for categorical attributes the interval the data is distributed on does not affect the value of the distance. It is strongly recommended to set clean_bool to True if also rescale_bool is set to true, since the rescaling operation is heavily affected by outliers.
N_SAMPLE : The S-RANK algorithm takes random samples of the dataset. This parameter is the number of samples to be taken. It is recommended to use at least 35.
SAMPLE_SIZE : The number of rows to include in each sample. It is recommended to use at least 0.25% of the dataset in each sample. If possible, tale samples of at least 1% of the total dataset.

Lastly, there are two ways to use the results:

ranker.info is a pandas.Dataframe containing the score obtained by each feature. The higher the score, the more important is the feature.
ranker.score is a list containing the features in order of importance. More important features first.

Reference

Dash, M., & Liu, H. (2000). Feature selection for clustering. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 1805, pp. 110-121). (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 1805). Springer Verlag.

Note:

The algorithm has been used is a small data science project and has given great results. However, i'm still testing its full capabilities.

Willinki / S-RANK