Repository files navigation

Mercury is a collaborative library that was developed by the Advanced Analytics community at BBVA. Originally, it was created as anInnerSource project but after some time, we decided to release certain parts of the project as Open Source.That's the case with themercury-robust package.

If you're interested in learning more about the Mercury project, we recommend reading this blogpost fromwww.bbvaaifactory.com

mercury-robust is a Python library designed for performing robust testing on machine learning models and datasets. It helps ensure that data workflows and models are robust against certain conditions, such as data drift, label leaking or the input data schema, by raising an exception when they fail. This library is intended for data scientists, machine learning engineers and anyone interested in ensuring the performance and robustness of their models in productive environments.

Errors or misbehaviours in machine learning models and datasets can have significant consequences, especially in sensitive domains such as healthcare or finance. It is important to ensure that model performance is align to what was measured in testing environemnts and robust to prevent harm to individuals or organizations that rely on them.mercury-robust helps ensure the robustness of machine learning models and datasets by providing a modular framework for performing tests.

mercury-robust provides two main types of tests:Data Tests andModel Tests. In addition, all tests can be added to a container class calledTestSuite

Data Tests receive a dataset as the main input argument and check different conditions. For example, the `CohortPerformanceTest checks whether some metrics perform poorly for some cohorts of data when compared to other groups. This is particularly relevant for measuring fairness in sensitive variables.

Model Tests involve data in combination with a machine learning model. For example, theModelSimplicityChecker evaluates if a simple baseline, trained in the same dataset, gives better or similar performance to a given model. It is used to check if added complexity contributes significantly to improve the model.

if the complexisty of the model is adecuate to the model measures the importance of every input feature and fails if the model has input features that add very marginal contribution.

This class provides an easy way to group tests and execute them together. Here's an example of aTestSuite that checks for input features that add very marginal importance to the model, the existence of linear combinations in those features, or some kind of data drift:

frommercury.robust.model_testsimportModelSimplicityCheckerfrommercury.robust.data_testsimportLinearCombinationsTest,DriftTestfrommercury.robust.suiteimportTestSuite# Create some testscomplexity_test=ModelSimplicityChecker(model=model,X_train=X_train,y_train=y_train,X_test=X_test,y_test=y_test,threshold=0.02,eval_fn=roc_auc_score)drift_test=DriftTest(df_test,train_schma,name="drift_train_test")lin_comb_test=LinearCombinationsTest(df_train)# Create the TestSuite with the teststest_suite=TestSuite(tests=[complexity_test,drift_test,lin_comb_test],run_safe=True)# Obtain resultstest_results=test_suite.get_results_as_df()

To help you get started with usingmercury-robust, we've created a cheatsheet that summarizes the main features and methods of the library. You can download the cheatsheet from here:RobustCheatsheet.pdf

The easiest way to installmercury-robust is usingpip:

pip install -U mercury-robust

This library is currently maintained by a dedicated team of data scientists and machine learning engineers from BBVA.

website:https://bbva.github.io/mercury-robust/site/

mercury.group@bbva.com