Practical techniques for interpreting machine learning models

Description

Transparency, auditability, and stability of predictive models and results are typically key differentiators in effective machine learning applications. Patrick Hall, Avni Wadhwa, and Mark Chan share tips and techniques learned through implementing interpretable machine learning solutions in industries like financial services, telecom, and health insurance. Using a set of publicly available and highly annotated examples, Patrick, Avni, and Mark teach several holistic approaches to interpretable machine learning. The examples use the well-known University of California Irvine (UCI) credit card dataset and popular open source packages to train constrained, interpretable machine learning models and visualize, explain, and test more complex machine learning models in the context of an example credit-risk application. Along the way, Patrick, Avni, and Mark draw on their applied experience to highlight crucial success factors and common pitfalls not typically discussed in blog posts and open source software documentation, such as the importance of both local and global interpretability and the approximate nature of nearly all machine learning interpretability techniques.

Outline:

Enhancing transparency in machine learning models with Python and XGBoost (example Jupyter notebook)

• Use monotonicity constraints to train an explainable—and potentially regulator-approvable—gradient boosting machine (GBM) credit risk model
• Use partial dependence plots and individual conditional expectation (ICE) plots to investigate the global and local mechanisms of the monotonic GBM and verify its monotonic behavior
• Use Shapley explanations to derive reason codes for model predictions.

Increasing transparency and accountability in your machine learning project with Python (example Jupyter notebook)

• Train a decision tree surrogate model on the original inputs and predictions of a complex GBM credit risk model to create an overall, approximate flowchart of the complex model’s predictions
• Compare the global variable importance from the GBM and from the surrogate decision tree and the interactions displayed in the decision tree with human domain expertise and reasonable expectations
• Use a variant of the leave-one-covariate-out (LOCO) technique to calculate the local contribution each input variable makes toward each model prediction, to enhance local understanding of the complex GBM’s behavior and the accountability of its predictions
• Rank local contributions to generate regulator-mandated reason codes that describe, in plain English, the GBM’s decision process for every prediction

Explaining your predictive models to business stakeholders with local interpretable model-agnostic explanations (LIME) using Python and H2O (example Jupyter Notebook)

• Explore a straightforward method of creating local samples for LIME that can be more appropriate for real-time scoring of new data in production applications
• Use LIME to understand local trends in the complex model’s predictions and calculate the local contribution of each input variable toward each model prediction
• Sort these contributions to create reason codes (i.e., regulator-mandated, plain English explanations of every model prediction)
• Validate LIME results to enhance trust in generated explanations using the local model’s R2 statistic and a ranked predictions plot

Testing machine learning models for accuracy, trustworthiness, and stability with Python and H2O (example Jupyter notebook)

• Explore sensitivity analysis—perhaps the most important validation technique for increasing trust in machine learning model predictions, because machine learning model predictions can vary drastically for small changes in input variable values, especially outside of training input domains
• Debug a trained GBM credit risk model using residual analysis to find problems arising from overfitting and outliers