feature-test 0.1.13

examines the relationships between a new feature and everyother feature in a dataset

feature_test

This package provides tools to test one feature against all other features in a dataset. It is intended to determine whether a new feature is a good candidate for addition to an established dataset used in a machine learning model. The benefit of this package is that instead of adding a feature to a dataset, running it through a long training and evaluation process, and then interpreting results, the feature can be quickly tested using feature_test; enabling users to quickly discard unuseful features and move forward with potentially impactful ones. This package provides the following tools to aid this determination.

• Correlation analysis
  • Calculate the correlations
  • Identify highly correlated feature
  • Categorize correlation
• Chi-Square tests
  • Calculate the chi-square statistic and p-vlaue
  • Categorize the chi-square result
• Recursive feature elimination (RFE)
  • Rank features by their importance
• Lasso regularization coefficients
  • Calculate the coefficients of a linear model using Lasso regularization
• Ridge regularization coefficients
  • Calculate the coefficients of a linear model using Ridge Regularization
• Decision tree coefficients
  • Calculate the coefficients for each feature using a decision tree model

• Special thanks to the AREN team for their guidance and Annie Tran for the open sourcing her feature test code.

Installation

python3 -m pip install feature-test

utils.Utils

class utils.Utils(X, columns)
A collection of functions that enables users to get data on a dataframe or adjust it for testing.

Parameters:

• X: pandas.DataFrame
  
  • A pandas.DataFrame containing the dataset
    
• columns: List
  
  • A list of strings correlating to features in a dataset.
    

Methods
_______________________________________________________________________________________

get_columns(X: pd.DataFrame)	Returns a list of column names in a dataframe.
exclude_columns(X: pd.DataFrame, columns: List)	Exclude a list of columns from a dataframe.

tests.Correlation

class tests.correlation(X, new_feature)

A suite of functions that calculates and reports the correlation coefficient between features and all other features in a dataset.

Parameters:

• X: pandas.DataFrame
  
  • A pandas.DataFrame containing the dataset
    
• new_feature: string
  
  • A string indicating the column in X to test against all other features
    

Methods
_______________________________________________________________________________________

calc_corr(X: pd.DataFrame, new_feature: str)	Returns a dataframe with the correlation values for each new_feature/feature combination.
similar_corr(X: pd.DataFrame, new_feature: str)	Returns a list of features highly correlated with the new_feature.
categorize_correlations(X: pd.DataFrame, correlation_threshold: float = 0.6)	Returns a dataframe with the correlations categorized. Possible values are high, medium, and low.
get_correlations(X: pd.DataFrame, new_feature: str)	Returns a dataframe of new_feature/feature combinations, their correlations, and their correlation category.

tests.ChiSquare

class tests.ChiSquare(X, new_feature)

Calculates the chi-squared statistic of the new feature against each categorical feature in a dataset. Also categorizes the chi-square result based on the p-value and effect size as measured by cramers v.

Parameters:

• X: pandas.DataFrame
  
  • A pandas.DataFrame containing the dataset
    
• new_feature: string
  
  • A string indicating the column in X to test against all other features
    

Methods
_______________________________________________________________________________________

calc_chi_sq(X: pd.DataFrame, new_feature: str)

Returns a dataframe that includes the chi-square result categorization.

tests.FeatureSelection

class tests.FeatureSelection(X, target)

Parameters:

• X: pandas.DataFrame
  
  • A pandas.DataFrame containing the dataset
    
• target: string
  
  • A string indicating the prediction column
    

Methods
_______________________________________________________________________________________

rfe_rankings(X: pd.DataFrame, target: str, classifier=None)	Returns a dataframe that includes the recursive feature elimination feature ranking.
lasso_rankings(X: pd.DataFrame, target: str)	Returns a dataframe that includes the linear model coefficients for features after lasso regularization.
ridge_coefficients(X: pd.DataFrame, target: str)	Returns a dataframe that includes the linear model coefficients for features after ridge regularization.
dtree_coefficients(X: pd.DataFrame, target: str)	Returns a dataframe that includes the decision tree model coefficients for features.
run_feature_classifiers(X: pd.DataFrame, target: str)	Returns a dataframe that includes the the results for rfe_rankings, lasso_rankings, ridge_rankings, and dtree_rankings.

Examples

---

from pandas import util

from feature_test.utils import Utils
from feature_test.feature_tests import Correlation, FeatureSelection, ChiSq

# Create a test dataset
df= util.testing.makeDataFrame()
df.head()

feature	A	B	C	D
lRhANYYD2r	0.572559	-1.409978	0.687618	-0.923502
YzYG07kY1O	0.145629	-1.446946	-0.003526	0.304385
cT3KK078Gt	-1.007378	1.263980	1.107897	0.844689
JW4Kg2EGVo	0.536701	-1.477372	-0.866873	1.539458
2mucO1cf2Z	-1.101875	0.518555	0.384916	-0.031403

c = Utils.get_columns(df)

['A', 'B', 'C', 'D']

corr_df = Correlation.calc_corr(df, 'A')
corr_df

feature_1	feature_2	corr
A	B	0.081662
A	C	0.203858
A	D	0.064999

rep_df = Correlation.get_correlations(df, 'D')
rep_df

feature_1	feature_2	corr	corr_cat
A	B	0.071466	low
A	C	0.105306	low
A	D	0.121130	low

ChiSq.calc_chi_sq(df, 'A')

feature_1	feature_2	chi_sq_cat
A	B	NOT SIGNIFICANT
A	C	NOT SIGNIFICANT
A	D	NOT SIGNIFICANT

FeatureSelection.rfe_rankings(df, 'A')

feature	rfe_rank
B	3
C	1
D	2

FeatureSelection.lasso_rankings(df, 'A')

feature	lasso_coef	lasso_importance
B	0.0	2.0
C	0.0	2.0
D	-0.0	2.0

FeatureSelection.ridge_rankings(df, 'A')

feature	ridge_coef	ridge_importance
B	0.050871	3.0
C	0.096362	1.0
D	-0.095220	2.0

FeatureSelection.dtree_rankings(df, 'A')

feature	random_forest_coefficient	random_forest_importance
B	0.323091	2.0
C	0.269673	3.0
D	0.407236	1.0

FeatureSelection.dtree_rankings(df, 'A')

feature	rfe_rank	lasso_coef	lasso_importance	ridge_coef	ridge_importance	random_forest_coefficient	random_forest_importance
B	3	0.0	2.0	0.050871	3.0	0.323091	2.0
C	1	0.0	2.0	0.096362	1.0	0.269673	3.0
D	2	-0.0	2.0	-0.095220	2.0	0.407236	1.0