8.6 Model Explainability

Use the OML4Py Explainability module to identify the important features that impact a trained model’s predictions.

Machine Learning Explainability (MLX) is the process of explaining and interpreting machine learning models. The OML MLX Python module supports the ability to help better understand a model's behavior and why it makes its predictions. MLX currently provides model-agnostic explanations for classification and regression tasks where explanations treat the ML model as a black-box, instead of using properties from the model to guide the explanation.

The global feature importance explainer object is the interface to the MLX permutation importance explainer. The global feature importance explainer identifies the most important features for a given model and data set. The explainer is model-agnostic and currently supports tabular classification and regression data sets with both numerical and categorical features.

The algorithm estimates feature importance by evaluating the model's sensitivity to changes in a specific feature. Higher sensitivity suggests that the model places higher importance on that feature when making its predictions than on another feature with lower sensitivity.

For information on the oml.GlobalFeatureImportance class attributes and methods, call help(oml.mlx.GlobalFeatureImportance) or see Oracle Machine Learning for Python API Reference.

Example 8-4 Binary Classification

This example uses the Breast Cancer binary classification data set. Load the data set into the database and a unique case id column.

import oml
from oml.mlx import GlobalFeatureImportance
import pandas as pd
import numpy as np
from sklearn import datasets

bc_ds = datasets.load_breast_cancer()
bc_data = bc_ds.data.astype(float)
X = pd.DataFrame(bc_data, columns=bc_ds.feature_names)
y = pd.DataFrame(bc_ds.target, columns=['TARGET'])
row_id = pd.DataFrame(np.arange(bc_data.shape[0]),
                      columns=['CASE_ID'])
df = oml.create(pd.concat([X, y, row_id], axis=1),
                table='BreastCancer')

Split the data set into train and test variables.

train, test = df.split(ratio=(0.8, 0.2), hash_cols='CASE_ID',
                       seed=32)
X, y = train.drop('TARGET'), train['TARGET']
X_test, y_test = test.drop('TARGET'), test['TARGET']

Train a Random Forest model.

model = oml.algo.rf(ODMS_RANDOM_SEED=32).fit(X, y, case_id='CASE_ID')
        "RF accuracy score = {:.2f}".format(model.score(X_test, y_test))

Create the MLX Global Feature Importance explainer, using the binary f1 metric.

gfi = GlobalFeatureImportance(mining_function='classification',
                              score_metric='f1', random_state=32,
                              parallel=4)

Run the explainer to generate the global feature importance. Here we construct an explanation using the train data set and then display the explanation.

explanation = gfi.explain(model, X, y, case_id='CASE_ID', n_iter=10)
explanation

Drop the BreastCancer table.

oml.drop('BreastCancer')

Listing for This Example

>>> import oml
>>> from oml.mlx import GlobalFeatureImportance
>>> import pandas as pd
>>> import numpy as np
>>> from sklearn import datasets
>>>
>>> bc_ds = datasets.load_breast_cancer()
>>> bc_data = bc_ds.data.astype(float)
>>> X = pd.DataFrame(bc_data, columns=bc_ds.feature_names)
>>> y = pd.DataFrame(bc_ds.target, columns=['TARGET'])
>>> row_id = pd.DataFrame(np.arange(bc_data.shape[0]),
...                       columns=['CASE_ID'])
>>> df = oml.create(pd.concat([X, y, row_id], axis=1), 
...                 table='BreastCancer')
>>>
>>> train, test = df.split(ratio=(0.8, 0.2), hash_cols='CASE_ID',
...                        seed=32)
>>> X, y = train.drop('TARGET'), train['TARGET']
>>> X_test, y_test = test.drop('TARGET'), test['TARGET']
>>>
>>> model = oml.algo.rf(ODMS_RANDOM_SEED=32).fit(X, y, case_id='CASE_ID')
...         "RF accuracy score = {:.2f}".format(model.score(X_test, y_test))
'RF accuracy score = 0.95'
>>>
>>> gfi = GlobalFeatureImportance(mining_function='classification', 
...                               score_metric='f1', random_state=32, 
...                               parallel=4)
>>>
>>> explanation = gfi.explain(model, X, y, case_id='CASE_ID', n_iter=10)
>>> explanation
Global Feature Importance:
[0] worst concave points: Value: 0.0263, Error: 0.0069
[1] worst perimeter: Value: 0.0077, Error: 0.0027
[2] worst radius: Value: 0.0076, Error: 0.0031
[3] worst area: Value: 0.0045, Error: 0.0037
[4] mean concave points: Value: 0.0034, Error: 0.0033
[5] worst texture: Value: 0.0017, Error: 0.0015
[6] area error: Value: 0.0012, Error: 0.0014
[7] worst concavity: Value: 0.0008, Error: 0.0008
[8] worst symmetry: Value: 0.0004, Error: 0.0007
[9] mean texture: Value: 0.0003, Error: 0.0007
[10] mean perimeter: Value: 0.0003, Error: 0.0015
[11] mean radius: Value: 0.0000, Error: 0.0000
[12] mean smoothness: Value: 0.0000, Error: 0.0000
[13] mean compactness: Value: 0.0000, Error: 0.0000
[14] mean concavity: Value: 0.0000, Error: 0.0000
[15] mean symmetry: Value: 0.0000, Error: 0.0000
[16] mean fractal dimension: Value: 0.0000, Error: 0.0000
[17] radius error: Value: 0.0000, Error: 0.0000
[18] texture error: Value: 0.0000, Error: 0.0000
[19] smoothness error: Value: 0.0000, Error: 0.0000
[20] compactness error: Value: 0.0000, Error: 0.0000
[21] concavity error: Value: 0.0000, Error: 0.0000
[22] concave points error: Value: 0.0000, Error: 0.0000
[23] symmetry error: Value: 0.0000, Error: 0.0000
[24] fractal dimension error: Value: 0.0000, Error: 0.0000
[25] worst compactness: Value: 0.0000, Error: 0.0000
[26] worst fractal dimension: Value: 0.0000, Error: 0.0000
[27] mean area: Value: -0.0001, Error: 0.0011
[28] worst smoothness: Value: -0.0003, Error: 0.0013

oml.drop('BreastCancer')

Example 8-5 Multi-Class Classification

This example uses the Iris multi-class classification data set. Load the data set into the database, adding a unique case id column.

import oml
from oml.mlx import GlobalFeatureImportance
import pandas as pd
import numpy as np
from sklearn import datasets

iris_ds = datasets.load_iris()
iris_data = iris_ds.data.astype(float)
X = pd.DataFrame(iris_data, columns=iris_ds.feature_names)
y = pd.DataFrame(iris_ds.target, columns=['TARGET'])
row_id = pd.DataFrame(np.arange(iris_data.shape[0]), 
                      columns=['CASE_ID'])
df = oml.create(pd.concat([X, y, row_id], axis=1), table='Iris')

Split the data set into train and test variables.

train, test = df.split(ratio=(0.8, 0.2), hash_cols='CASE_ID',
                       seed=32)
X, y = train.drop('TARGET'), train['TARGET']
X_test, y_test = test.drop('TARGET'), test['TARGET']

Train an SVM model.

model = oml.algo.svm(ODMS_RANDOM_SEED=32).fit(X, y, case_id='CASE_ID')
"SVM accuracy score = {:.2f}".format(model.score(X_test, y_test))

Create the MLX Global Feature Importance explainer, using the f1_weighted metric.

gfi = GlobalFeatureImportance(mining_function='classification', 
                              score_metric='f1_weighted', 
                              random_state=32, parallel=4)

Run the explainer to generate the global feature importance. Here, we use the test data set. Display the explanation.

explanation = gfi.explain(model, X_test, y_test,
                          case_id='CASE_ID', n_iter=10)
explanation

Drop the Iris table.

oml.drop('Iris')

Listing for This Example

>>> import oml
>>> from oml.mlx import GlobalFeatureImportance
>>> import pandas as pd
>>> import numpy as np
>>> from sklearn import datasets
>>>
>>> iris_ds = datasets.load_iris()
>>> iris_data = iris_ds.data.astype(float)
>>> X = pd.DataFrame(iris_data, columns=iris_ds.feature_names)
>>> y = pd.DataFrame(iris_ds.target, columns=['TARGET'])
>>> row_id = pd.DataFrame(np.arange(iris_data.shape[0]),
...                       columns=['CASE_ID'])
>>> df = oml.create(pd.concat([X, y, row_id], axis=1), table='Iris')
>>>
>>> train, test = df.split(ratio=(0.8, 0.2), hash_cols='CASE_ID',
...                        seed=32)
>>> X, y = train.drop('TARGET'), train['TARGET']
>>> X_test, y_test = test.drop('TARGET'), test['TARGET']
>>>
>>> model = oml.algo.svm(ODMS_RANDOM_SEED=32).fit(X, y, case_id='CASE_ID')
>>> "SVM accuracy score = {:.2f}".format(model.score(X_test, y_test))
'SVM accuracy score = 0.94'
>>>
>>> gfi = GlobalFeatureImportance(mining_function='classification', 
...                               score_metric='f1_weighted', 
...                               random_state=32, parallel=4)
>>>
>>> explanation = gfi.explain(model, X_test, y_test, 
...                           case_id='CASE_ID', n_iter=10)
>>> explanation
Global Feature Importance:
[0] petal length (cm): Value: 0.3462, Error: 0.0824
[1] petal width (cm): Value: 0.2417, Error: 0.0687
[2] sepal width (cm): Value: 0.0926, Error: 0.0452
[3] sepal length (cm): Value: 0.0253, Error: 0.0152

>>> oml.drop('Iris')

Example 8-6 Regression

This example uses the Boston regression data set. Load the data set into the database, adding a unique case id column.

import oml
from oml.mlx import GlobalFeatureImportance
import pandas as pd
import numpy as np
from sklearn import datasets

boston_ds = datasets.load_boston()
boston_data = boston_ds.data
X = pd.DataFrame(boston_data, columns=boston_ds.feature_names)
y = pd.DataFrame(boston_ds.target, columns=['TARGET'])
row_id = pd.DataFrame(np.arange(boston_data.shape[0]),
                      columns=['CASE_ID'])
df = oml.create(pd.concat([X, y, row_id], axis=1), table='Boston')

Split the data set into train and test variables.

train, test = df.split(ratio=(0.8, 0.2), hash_cols='CASE_ID', seed=32)
X, y = train.drop('TARGET'), train['TARGET']
X_test, y_test = test.drop('TARGET'), test['TARGET']

Train a Neural Network regression model.

model = oml.algo.nn(mining_function='regression', 
                    ODMS_RANDOM_SEED=32).fit(X, y, case_id='CASE_ID')
"NN R^2 score = {:.2f}".format(model.score(X_test, y_test))

Create the MLX Global Feature Importance explainer, using the r2 metric.

gfi = GlobalFeatureImportance(mining_function='regression', 
                              score_metric='r2', random_state=32, 
                              parallel=4)

Run the explainer to generate the global feature importance. Here, we use the test data set. Display the explanation.

explanation = gfi.explain(model, df, 'TARGET',
                          case_id='CASE_ID', n_iter=10)
explanation

Drop the Boston table.

oml.drop('Boston')

Listing for This Example

>>> import oml
>>> from oml.mlx import GlobalFeatureImportance
>>> import pandas as pd
>>> import numpy as np
>>> from sklearn import datasets
>>>
>>> boston_ds = datasets.load_boston()
>>> boston_data = boston_ds.data
>>> X = pd.DataFrame(boston_data, columns=boston_ds.feature_names)
>>> y = pd.DataFrame(boston_ds.target, columns=['TARGET'])
>>> row_id = pd.DataFrame(np.arange(boston_data.shape[0]),
...                       columns=['CASE_ID'])
>>> df = oml.create(pd.concat([X, y, row_id], axis=1), table='Boston')
>>>
>>> train, test = df.split(ratio=(0.8, 0.2), hash_cols='CASE_ID',
...                        seed=32)
>>> X, y = train.drop('TARGET'), train['TARGET']
>>> X_test, y_test = test.drop('TARGET'), test['TARGET']
>>>
>>> model = oml.algo.nn(mining_function='regression', 
..                      ODMS_RANDOM_SEED=32).fit(X, y, case_id='CASE_ID')
>>> "NN R^2 score = {:.2f}".format(model.score(X_test, y_test))
'NN R^2 score = 0.85'
>>>
>>> gfi = GlobalFeatureImportance(mining_function='regression', 
...                               score_metric='r2', random_state=32, 
...                               parallel=4)
>>>
>>> explanation = gfi.explain(model, df, 'TARGET', 
...                           case_id='CASE_ID', n_iter=10)
>>> explanation
Global Feature Importance:
[0] LSTAT: Value: 0.7686, Error: 0.0513
[1] RM: Value: 0.5734, Error: 0.0475
[2] CRIM: Value: 0.5131, Error: 0.0345
[3] DIS: Value: 0.4170, Error: 0.0632
[4] NOX: Value: 0.2592, Error: 0.0206
[5] AGE: Value: 0.2083, Error: 0.0212
[6] RAD: Value: 0.1956, Error: 0.0188
[7] INDUS: Value: 0.1792, Error: 0.0199
[8] B: Value: 0.0982, Error: 0.0146
[9] PTRATIO: Value: 0.0822, Error: 0.0069
[10] TAX: Value: 0.0566, Error: 0.0139
[11] ZN: Value: 0.0397, Error: 0.0081
[12] CHAS: Value: 0.0125, Error: 0.0045

>>> oml.drop('Boston')