InterpretML: A Unified Framework for Machine Learning Interpretability (1909.09223v1)

Abstract: InterpretML is an open-source Python package which exposes machine learning interpretability algorithms to practitioners and researchers. InterpretML exposes two types of interpretability - glassbox models, which are machine learning models designed for interpretability (ex: linear models, rule lists, generalized additive models), and blackbox explainability techniques for explaining existing systems (ex: Partial Dependence, LIME). The package enables practitioners to easily compare interpretability algorithms by exposing multiple methods under a unified API, and by having a built-in, extensible visualization platform. InterpretML also includes the first implementation of the Explainable Boosting Machine, a powerful, interpretable, glassbox model that can be as accurate as many blackbox models. The MIT licensed source code can be downloaded from github.com/microsoft/interpret.

• The paper presents a unified framework integrating glassbox models with blackbox explainability techniques to provide comprehensive interpretability in ML.
• It introduces the Explainable Boosting Machine (EBM), a novel glassbox model that rivals traditional methods on AUROC while ensuring transparency.
• The framework emphasizes modularity and compatibility, enabling easy comparison and customization for critical applications like healthcare and finance.

InterpretML: A Unified Framework for Machine Learning Interpretability

The paper, "InterpretML: A Unified Framework for Machine Learning Interpretability," presents an open-source Python package aimed at enhancing the interpretability of machine learning models. Integrating both glassbox models and blackbox explainability techniques, InterpretML serves as a comprehensive tool for researchers and practitioners seeking to understand the underlying mechanisms of various machine learning methodologies.

Interpretability Framework

InterpretML distinguishes between two primary interpretability paradigms: glassbox models and blackbox explainability techniques. Glassbox models, such as linear models, rule lists, and generalized additive models (GAMs), are explicitly designed to be interpretable. On the other hand, blackbox techniques like Partial Dependence and LIME provide post-hoc explanations for otherwise opaque models. This dual approach allows users to not only select inherently interpretable models but also to derive insights from complex models that are traditionally difficult to interpret.

Design Principles

The design of InterpretML is driven by four key principles:

1. Ease of Comparison: A unified API, consistent with scikit-learn, enables straightforward comparison between different interpretability algorithms.
2. Fidelity: The package endeavors to faithfully reproduce original algorithm designs and visualizations.
3. Compatibility: Leveraging existing open-source tools, it maintains strong interoperability with popular projects such as Jupyter Notebook and libraries like plotly and SALib.
4. Modularity: Users can extend or utilize individual components of the framework without importing the entire package, enabling scalability and customization.

Explainable Boosting Machine

A notable contribution of InterpretML is the Explainable Boosting Machine (EBM), a novel algorithmic innovation. EBM stands out as a glassbox model that rivals blackbox models like Random Forest or Boosted Trees in accuracy. Conceptually, EBM is a generalized additive model enhanced by contemporary machine learning tactics such as bagging and gradient boosting. Specifically, it operates by learning feature functions $f_j(x_j)$ iteratively in a round-robin manner, ensuring superior interpretability by mitigating feature co-linearity. Additionally, EBM incorporates pairwise interaction terms, allowing it to preserve intelligibility while boosting predictive performance.

The paper provides comprehensive performance evaluation across various datasets, demonstrating that EBM frequently outperforms traditional models such as Logistic Regression, Random Forest, and XGBoost in terms of AUROC. Despite the predictive competitiveness, EBM incurs higher training costs due to its structured learning constraints. However, EBM compensates with fast prediction times and minimal memory usage, making it a viable option for deployment in production.

Implications and Future Directions

InterpretML addresses a significant gap in the accessibility of interpretability techniques by consolidating them into a unified platform. The framework encourages the adoption of interpretable models in critical domains such as healthcare, finance, and judicial systems, where understanding model behavior is paramount. The introduction of the Explainable Boosting Machine expands the possibilities for practitioners seeking models that balance accuracy with transparency.

Future work in this domain may include refining EBM's training efficiency and extending the framework's compatibility with emerging interpretability methodologies. The evolution of InterpretML could catalyze further advancements in creating models that are not only performant but also accountable, fostering trust and reliability in AI systems.