Binary Classifier Class

Binary classification model which wraps python scikit-learn's Regularized Logistic Regression with series of specific functionalities.

All input features X are assumed to be given as pandas DataFrames with a mix of numeric and categorical data. The output y is assumed to be in form of a numpy array of 0-s and 1-s.

X = pd.DataFrame({'v1': ['c', 'b', 'a'], 'v2': [2, 6, 4]}) 
y = np.array([0, 0, 1])

The package is organized in following folders:

• src: Main class codes and utilities
• test: Test cases for unittest
• docs: Documentation and questions
• data: Test data and exploratory data analysis

There are two more classes developed beside the main classifier classes as below which are hosted under robo_prep file.

• RoboImputer: Filling up the missing values
• RoboFeaturizer: generalized pre-processing of input data for classifier class

Imputer

The class RoboImputer performs filling up the missing values with the laws the simple imputing rules below. This class is used in pro-processing of data thorough the RoboFeaturizer class.

• Column dtype : Imputing Value
• object/category : most frequent value in the column
• float : mean of column values
• other dtype : median of column

Initialization syntax:

i = RoboImputer()

Featurizer

All the generalized pre-processing of input data for classifier class are taken care of using the RoboFeaturizer class which also gets helped in imputing missing values by the auxiliary RoboImputer class above.

Initialization syntax:

f = RoboFeaturizer(
             max_unique_for_discrete=10,
             max_missing_to_keep=0.80,
             add_missing_flag=False,
             encode_categorical=True,
             max_category_for_ohe=10,
             scaler = StandardScaler()
             )

where

• max_unique_for_discrete [int]: Encode columns of up to this threshold of unique numeric values, using one-hot-encoding.
• max_missing_to_keep [0=<float<=1]: Drop columns if fraction (%) of missing values are higher than this threshold
• add_missing_flag [Bool]: For columns with missing values smaller than the threshold, add a binary column to mark missing data points.
• encode_categorical [Bool]: Encode categorical (non-numerical) columns
• max_category_for_ohe [int]: Encode categorical columns with number of categories up to this threshold, use one-hot-encoding and for the rest use alternate. Only effective if encode_categorical=True
• scaler [None or sklearn scaler]: No scaling of input if None, scale input using 'fit_transform' method

Classifier based on Regularize Logistic Regression

The class follows standard scikit-learn template in initialization and method definition. Custom-defined methods for this class are listed below:

Initialization syntax:

f = RoboLogistic(
             max_unique_for_discrete=10,
             max_missing_to_keep=0.80,
             add_missing_flag=False,
             encode_categorical=True,
             max_category_for_ohe=10,
             scaler = StandardScaler(),
             ...)

where the initialization parameters are combination of RoboFeaturizer class discussed above and scikit-learn's LogisticRegression class.

fit:

Fit on training data.

self.fit(X, y)

predict:

Predict class labels on new data.

self.predict(X)

predict_proba:

Predict the probability of each label.

self.predict_proba(X)

evaluate:

Find the value of the F1-score, LogLoss metric.

self.evaluate(X, y)

tune_parameters:

Run K-fold cross validation to choose the best parameters given k as number of folds and dictionary of hyper-parameters for search.

self.tune_parameters(X, y, k=5, hypeparam_grid=None)

Unit Test

Exploratory data analysis of test dataset was performed and provided for reference in
a notebook.

For the purpose of unit test, a simple class of TestData was developed to load either the bad credit dataset or the breast cancer dataset from scikit-learn. The class simply loads the data and helps with generating X and y Unit test was developed with a focus on each of the three classes defined and discussed above.

TestData Class

X, y = TestData().make_X_y()

Test Cases

For each of the classes:

• Check if proper validation and error is provided if a class instance is not fit and called for prediction or transform.
• Check if input is in the correct shape during calling transform or prediction.
• Check class-specific cases to ensure class is behaving according to the design.

There is room for more test cases and integration with this package. Coommand below can be used from command line to start testing:

$ python -m unittest discover

Tests on Imputer Class

• test_bad_input_shape
• test_unfit_instance
• test_filevalue1
• test_filevalue2

Tests on Featurizer Class

• test_bad_input_shape
• test_unfit_instance
• test_missing_value
• test_new_category

Tests on Logistic Class

• test_bad_input_shape
• test_unfit_instance
• test_result_formats
• tesT_reproduce