Maximizing the Interpretation of Black-Box Models

Supplementary Materials for CONVENTION A | ASIA

Overview

Welcome to the supplementary repository for our presentation Maximizing the Interpretation of Black-Box Models at CONVENTION A | ASIA.

This project demonstrates the application of Maximum Interpretation Decomposition (MID), a novel approach in Interpretable Machine Learning (IML) designed to bridge the gap between high-performance “black-box” models and the transparency requirements of actuarial practice.

Key Topics of the Presentation

  • Theoretical Foundation: Understanding the decomposition logic of complex models into interpretable parts: intercept, first-order main effects, and second-order interaction effects by MID.
  • Actuarial Application: Benchmarking MID using the industry-standard freMTPL2freq dataset using open-source software: {midr}, {midnight} (R), and {midlearn} (Python).

Project Structure

This Quarto website contains the following sections:

  1. Dataset: Data cleaning and feature engineering of the French Motor Third-Party Liability dataset.
  2. R Demo: A demonstration of the {midr} and {midlearn} packages in R.
  3. Python Demo: A demonstration of the {midlearn} library, providing a seamless interface for Python users.
  4. Going Beyond: A demonstration of a new feature in {midr} 0.6.0: interpretation of survival models.

Software

  • {midr}: GitHub / CRAN
    R package for Maximum Interpretation Decomposition in R
  • {midnight}: GitHub
    R package to integrate {midr} to the {tidymodels} ecosystem
  • {midlearn}: GitHub / PyPI
    Python library to integrate {midr} to the {scikit-learn} ecosystem

Acknowledgment & Reference

  • {CASdatasets}: CRAN
    This demonstration utilizes the freMTPL2freq dataset from the {CASdatasets} package.

  • AITools4Actuaries: Website
    We would like to acknowledge the project for their foundational work on benchmarking ML models in insurance, which served as a reference, especially for our data preprocessing pipeline.