Original Feature Importance Explainer for MOJO Models (Naive Shapley method) Demo
This example demonstrates how to interpret a Driverless AI MOJO model using the H2O Eval Studio library and retrieve the data and plot with original features importances.
[1]:
import os
import logging
import datatable
import daimojo
import webbrowser
from h2o_sonar import interpret
from h2o_sonar.lib.api import commons
from h2o_sonar.lib.api import explainers
from h2o_sonar.explainers import fi_naive_shapley_explainer as explainer
from h2o_sonar.lib.api.models import ModelApi
[2]:
# explainer description
interpret.describe_explainer(explainer.NaiveShapleyMojoFeatureImportanceExplainer)
[2]:
{'id': 'h2o_sonar.explainers.fi_naive_shapley_explainer.NaiveShapleyMojoFeatureImportanceExplainer',
'name': 'NaiveShapleyMojoFeatureImportanceExplainer',
'display_name': 'Shapley Values for Original Features of MOJO Models (Naive Method)',
'description': 'Shapley values for original features of (Driverless AI) MOJO models are approximated from the accompanying Shapley values for transformed features with the Naive Shapley method. This method makes the assumption that input features to a transformer are independent. For example, if the transformed feature, feature1_feature2, has a Shapley value of 0.5, then the Shapley value of the original features feature1 and feature2 will be 0.25 each.',
'model_types': ['iid'],
'can_explain': ['regression', 'binomial', 'multinomial'],
'explanation_scopes': ['global_scope', 'local_scope'],
'explanations': [{'explanation_type': 'global-feature-importance',
'name': 'GlobalFeatImpExplanation',
'category': None,
'scope': 'global',
'has_local': None,
'formats': []},
{'explanation_type': 'local-feature-importance',
'name': 'LocalFeatImpExplanation',
'category': None,
'scope': 'local',
'has_local': None,
'formats': []}],
'parameters': [{'name': 'sample_size',
'description': 'Sample size.',
'comment': '',
'type': 'int',
'val': 100000,
'predefined': [],
'tags': [],
'min_': 0.0,
'max_': 0.0,
'category': ''},
{'name': 'fast_approx_contribs',
'description': 'Speed up predictions with fast contributions predictions approximation.',
'comment': '',
'type': 'bool',
'val': True,
'predefined': [],
'tags': [],
'min_': 0.0,
'max_': 0.0,
'category': ''},
{'name': 'leakage_warning_threshold',
'description': 'The threshold above which to report a potentially detected feature importance leak problem.',
'comment': '',
'type': 'float',
'val': 0.95,
'predefined': [],
'tags': [],
'min_': 0.0,
'max_': 0.0,
'category': ''}],
'keywords': ['run-by-default',
'explains-original-feature-importance',
'is_fast',
'h2o-sonar']}
Interpretation
[10]:
# dataset
dataset_path = "../../data/creditcard.csv"
target_col = "default payment next month"
# model
mojo_path = "../../data/models/creditcard-binomial.mojo"
mojo_model = daimojo.model(mojo_path)
model = ModelApi().create_model(
model_src=mojo_model,
target_col=target_col,
used_features=list(mojo_model.feature_names),
)
# results
results_location = "./results"
os.makedirs(results_location, exist_ok=True)
[11]:
interpretation = interpret.run_interpretation(
dataset=dataset_path,
model=model,
target_col=target_col,
results_location=results_location,
explainers=[explainer.NaiveShapleyMojoFeatureImportanceExplainer.explainer_id()],
log_level=logging.INFO,
)
h2o_sonar.explainers.fi_naive_shapley_explainer.NaiveShapleyMojoFeatureImportanceExplainer: progress 20.0%
h2o_sonar.explainers.fi_naive_shapley_explainer.NaiveShapleyMojoFeatureImportanceExplainer: progress 90.0%
h2o_sonar.explainers.fi_naive_shapley_explainer.NaiveShapleyMojoFeatureImportanceExplainer: progress 90.0%
Explainer Result
[12]:
# retrieve the result
result = interpretation.get_explainer_result(
explainer.NaiveShapleyMojoFeatureImportanceExplainer.explainer_id()
)
[13]:
# open interpretation HTML report in web browser
webbrowser.open(interpretation.result.get_html_report_location())
[13]:
True
[14]:
# summary
result.summary()
[14]:
{'id': 'h2o_sonar.explainers.fi_naive_shapley_explainer.NaiveShapleyMojoFeatureImportanceExplainer',
'name': 'NaiveShapleyMojoFeatureImportanceExplainer',
'display_name': 'Shapley Values for Original Features of MOJO Models (Naive Method)',
'description': 'Shapley values for original features of (Driverless AI) MOJO models are approximated from the accompanying Shapley values for transformed features with the Naive Shapley method. This method makes the assumption that input features to a transformer are independent. For example, if the transformed feature, feature1_feature2, has a Shapley value of 0.5, then the Shapley value of the original features feature1 and feature2 will be 0.25 each.',
'model_types': ['iid'],
'can_explain': ['regression', 'binomial', 'multinomial'],
'explanation_scopes': ['global_scope', 'local_scope'],
'explanations': [{'explanation_type': 'global-feature-importance',
'name': 'Shapley Values for Original Features of MOJO Models (Naive Method)',
'category': 'DAI MODEL',
'scope': 'global',
'has_local': 'local-feature-importance',
'formats': ['application/vnd.h2oai.json+datatable.jay',
'application/vnd.h2oai.json+csv',
'application/json']},
{'explanation_type': 'local-feature-importance',
'name': 'Shapley Values for Original Features of MOJO Models (Naive Method)',
'category': 'DAI MODEL',
'scope': 'local',
'has_local': None,
'formats': ['application/vnd.h2oai.json+datatable.jay']},
{'explanation_type': 'global-html-fragment',
'name': 'Shapley Values for Original Features of MOJO Models (Naive Method)',
'category': 'DAI MODEL',
'scope': 'global',
'has_local': None,
'formats': ['text/html']}],
'parameters': [{'name': 'sample_size',
'description': 'Sample size.',
'comment': '',
'type': 'int',
'val': 100000,
'predefined': [],
'tags': [],
'min_': 0.0,
'max_': 0.0,
'category': ''},
{'name': 'fast_approx_contribs',
'description': 'Speed up predictions with fast contributions predictions approximation.',
'comment': '',
'type': 'bool',
'val': True,
'predefined': [],
'tags': [],
'min_': 0.0,
'max_': 0.0,
'category': ''}],
'keywords': ['run-by-default',
'explains-original-feature-importance',
'is_fast',
'h2o-sonar']}
[15]:
# parameters
result.params()
[15]:
{'sample_size': 100000, 'fast_approx_contribs': True}
Display Data
[16]:
result.data()
[16]:
| feature | importance | |
|---|---|---|
| ▪▪▪▪ | ▪▪▪▪▪▪▪▪ | |
| 0 | PAY_0 | 0.484314 |
| 1 | PAY_2 | 0.177288 |
| 2 | LIMIT_BAL | 0.14293 |
| 3 | PAY_AMT4 | 0.125988 |
| 4 | PAY_AMT2 | 0.109839 |
| 5 | BILL_AMT1 | 0.0856685 |
| 6 | PAY_3 | 0.0460469 |
| 7 | PAY_AMT3 | 0.0405643 |
| 8 | PAY_6 | 0.035129 |
| 9 | BILL_AMT2 | 0.0307533 |
| 10 | PAY_4 | 0.0292232 |
| 11 | BILL_AMT6 | 0.0230935 |
| 12 | PAY_5 | 0.0203799 |
| 13 | PAY_AMT1 | 0.014774 |
| 14 | EDUCATION | 0.00983341 |
| 15 | AGE | 0.00831316 |
| 16 | MARRIAGE | 0.00725285 |
| 17 | PAY_AMT6 | 0.00588243 |
| 18 | PAY_AMT5 | 0.00541516 |
| 19 | BILL_AMT5 | 0.00464642 |
| 20 | BILL_AMT4 | 0.00144066 |
| 21 | BILL_AMT3 | 0.000818275 |
Plot Feature Importance Data
[17]:
result.plot()
Save Explainer Log and Data
[18]:
# save the explainer log
log_file_path = "./feature-importance-demo.log"
result.log(path=log_file_path)
[19]:
!cat $log_file_path
[20]:
# save the explainer data
result.zip(file_path="./feature-importance-demo-archive.zip")
[21]:
!unzip -l feature-importance-demo-archive.zip
Archive: feature-importance-demo-archive.zip
Length Date Time Name
--------- ---------- ----- ----
2991 2023-03-12 23:02 explainer_h2o_sonar_explainers_fi_naive_shapley_explainer_NaiveShapleyMojoFeatureImportanceExplainer_6b2f1f72-b743-4ffd-84ec-b2f8383153ca/result_descriptor.json
1762120 2023-03-12 23:02 explainer_h2o_sonar_explainers_fi_naive_shapley_explainer_NaiveShapleyMojoFeatureImportanceExplainer_6b2f1f72-b743-4ffd-84ec-b2f8383153ca/work/shapley.orig.feat.bin
1742542 2023-03-12 23:02 explainer_h2o_sonar_explainers_fi_naive_shapley_explainer_NaiveShapleyMojoFeatureImportanceExplainer_6b2f1f72-b743-4ffd-84ec-b2f8383153ca/work/shapley_formatted_orig_feat.zip
4693127 2023-03-12 23:02 explainer_h2o_sonar_explainers_fi_naive_shapley_explainer_NaiveShapleyMojoFeatureImportanceExplainer_6b2f1f72-b743-4ffd-84ec-b2f8383153ca/work/shapley.orig.feat.csv
40216 2023-03-12 23:02 explainer_h2o_sonar_explainers_fi_naive_shapley_explainer_NaiveShapleyMojoFeatureImportanceExplainer_6b2f1f72-b743-4ffd-84ec-b2f8383153ca/work/y_hat.bin
110 2023-03-12 23:02 explainer_h2o_sonar_explainers_fi_naive_shapley_explainer_NaiveShapleyMojoFeatureImportanceExplainer_6b2f1f72-b743-4ffd-84ec-b2f8383153ca/global_html_fragment/text_html.meta
378 2023-03-12 23:02 explainer_h2o_sonar_explainers_fi_naive_shapley_explainer_NaiveShapleyMojoFeatureImportanceExplainer_6b2f1f72-b743-4ffd-84ec-b2f8383153ca/global_html_fragment/text_html/explanation.html
24441 2023-03-12 23:02 explainer_h2o_sonar_explainers_fi_naive_shapley_explainer_NaiveShapleyMojoFeatureImportanceExplainer_6b2f1f72-b743-4ffd-84ec-b2f8383153ca/global_html_fragment/text_html/fi-class-0.png
201 2023-03-12 23:02 explainer_h2o_sonar_explainers_fi_naive_shapley_explainer_NaiveShapleyMojoFeatureImportanceExplainer_6b2f1f72-b743-4ffd-84ec-b2f8383153ca/local_feature_importance/application_vnd_h2oai_json_datatable_jay.meta
1762120 2023-03-12 23:02 explainer_h2o_sonar_explainers_fi_naive_shapley_explainer_NaiveShapleyMojoFeatureImportanceExplainer_6b2f1f72-b743-4ffd-84ec-b2f8383153ca/local_feature_importance/application_vnd_h2oai_json_datatable_jay/feature_importance_class_0.jay
831 2023-03-12 23:02 explainer_h2o_sonar_explainers_fi_naive_shapley_explainer_NaiveShapleyMojoFeatureImportanceExplainer_6b2f1f72-b743-4ffd-84ec-b2f8383153ca/local_feature_importance/application_vnd_h2oai_json_datatable_jay/explanation.json
40216 2023-03-12 23:02 explainer_h2o_sonar_explainers_fi_naive_shapley_explainer_NaiveShapleyMojoFeatureImportanceExplainer_6b2f1f72-b743-4ffd-84ec-b2f8383153ca/local_feature_importance/application_vnd_h2oai_json_datatable_jay/y_hat.bin
2 2023-03-12 23:02 explainer_h2o_sonar_explainers_fi_naive_shapley_explainer_NaiveShapleyMojoFeatureImportanceExplainer_6b2f1f72-b743-4ffd-84ec-b2f8383153ca/model_problems/problems_and_actions.json
0 2023-03-12 23:02 explainer_h2o_sonar_explainers_fi_naive_shapley_explainer_NaiveShapleyMojoFeatureImportanceExplainer_6b2f1f72-b743-4ffd-84ec-b2f8383153ca/log/explainer_run_6b2f1f72-b743-4ffd-84ec-b2f8383153ca.log
143 2023-03-12 23:02 explainer_h2o_sonar_explainers_fi_naive_shapley_explainer_NaiveShapleyMojoFeatureImportanceExplainer_6b2f1f72-b743-4ffd-84ec-b2f8383153ca/global_feature_importance/application_json.meta
163 2023-03-12 23:02 explainer_h2o_sonar_explainers_fi_naive_shapley_explainer_NaiveShapleyMojoFeatureImportanceExplainer_6b2f1f72-b743-4ffd-84ec-b2f8383153ca/global_feature_importance/application_vnd_h2oai_json_csv.meta
185 2023-03-12 23:02 explainer_h2o_sonar_explainers_fi_naive_shapley_explainer_NaiveShapleyMojoFeatureImportanceExplainer_6b2f1f72-b743-4ffd-84ec-b2f8383153ca/global_feature_importance/application_vnd_h2oai_json_datatable_jay.meta
880 2023-03-12 23:02 explainer_h2o_sonar_explainers_fi_naive_shapley_explainer_NaiveShapleyMojoFeatureImportanceExplainer_6b2f1f72-b743-4ffd-84ec-b2f8383153ca/global_feature_importance/application_vnd_h2oai_json_datatable_jay/feature_importance_class_0.jay
1245 2023-03-12 23:02 explainer_h2o_sonar_explainers_fi_naive_shapley_explainer_NaiveShapleyMojoFeatureImportanceExplainer_6b2f1f72-b743-4ffd-84ec-b2f8383153ca/global_feature_importance/application_vnd_h2oai_json_datatable_jay/explanation.json
1593 2023-03-12 23:02 explainer_h2o_sonar_explainers_fi_naive_shapley_explainer_NaiveShapleyMojoFeatureImportanceExplainer_6b2f1f72-b743-4ffd-84ec-b2f8383153ca/global_feature_importance/application_json/feature_importance_class_0.json
578 2023-03-12 23:02 explainer_h2o_sonar_explainers_fi_naive_shapley_explainer_NaiveShapleyMojoFeatureImportanceExplainer_6b2f1f72-b743-4ffd-84ec-b2f8383153ca/global_feature_importance/application_json/explanation.json
726 2023-03-12 23:02 explainer_h2o_sonar_explainers_fi_naive_shapley_explainer_NaiveShapleyMojoFeatureImportanceExplainer_6b2f1f72-b743-4ffd-84ec-b2f8383153ca/global_feature_importance/application_vnd_h2oai_json_csv/feature_importance_class_0.csv
577 2023-03-12 23:02 explainer_h2o_sonar_explainers_fi_naive_shapley_explainer_NaiveShapleyMojoFeatureImportanceExplainer_6b2f1f72-b743-4ffd-84ec-b2f8383153ca/global_feature_importance/application_vnd_h2oai_json_csv/explanation.json
--------- -------
10075385 23 files
[ ]: