shapr: Explaining Machine Learning Models with Conditional Shapley Values in R and Python

TL;DR

The paper presents shapr, an R package (and shaprpy for Python) that provides conditional Shapley value explanations for a wide range of predictive models. It advances model interpretability by emphasizing conditional (distribution-aware) explanations, implements a comprehensive set of estimation approaches (e.g., independence, Gaussian, copula, ctree, VA EAC, regression-based), and offers iterative convergence checks, parallelization, and rich visualizations. It also covers extensions to asymmetric and causal Shapley values, supports forecasting models via explain_forecast(), and includes a Python wrapper to bring these capabilities to Python workflows. The work enables accurate, scalable, and flexible explanations in both tabular and time-series contexts, with practical guidance for method selection via the MSE_v criterion and extensive examples on real datasets.

Abstract

This paper introduces the shapr R package, a versatile tool for generating Shapley value based prediction explanations for machine learning and statistical regression models. Moreover, the shaprpy Python library brings the core capabilities of shapr to the Python ecosystem. Shapley values originate from cooperative game theory in the 1950s, but have over the past few years become a widely used method for quantifying how a model's features/covariates contribute to specific prediction outcomes. The shapr package emphasizes conditional Shapley value estimates, providing a comprehensive range of approaches for accurately capturing feature dependencies -- a crucial aspect for correct model explanation, typically lacking in similar software. In addition to regular tabular data, the shapr R package includes specialized functionality for explaining time series forecasts. The package offers a minimal set of user functions with sensible default values for most use cases while providing extensive flexibility for advanced users to fine-tune computations. Additional features include parallelized computations, iterative estimation with convergence detection, and rich visualization tools. shapr also extends its functionality to compute causal and asymmetric Shapley values when causal information is available. Overall, the shapr and shaprpy packages aim to enhance the interpretability of predictive models within a powerful and user-friendly framework.

Figures (6)

• Figure 1: Schematic overview of the available approaches in shapr for computing conditional Shapley value explanations. Parsnip* indicates that the regression-based approaches can use any regression methods available in the parsnip package Parsnip, including user-specified methods. The shapr package also implements the timeseries approach, which is specifically designed for time series data and not applicable to regular tabular datasets. See jullum2021groupshapley for details.
• Figure 2: The output to the console when creating exp_iter_ctree with explain() using verbose = c("basic", "convergence").
• Figure 3: Scatter plot of the Shapley values for the two features atemp and windspeed, created by shapr.plot. The plot shows that the Shapley values are largest for atemp around 20-30 and a small windspeed, meaning that in these situations, the temperature and windspeed increases the predicted number of bike rentals the most. For very low atemp and high windspeed, the predicted number of bike rentals are severely reduced. These effects seems very natural.
• Figure 4: Illustration of a waterfall plot of a single explicand with group-wise Shapley values created by shapr.plot. The waterfall plot illustrates how the cumulative contributions of feature groups sum to the prediction, starting from the mean prediction/response $\phi_0$ and adding contributions in order of increasing absolute value. For this explicand, the weather group increases the predicted number of bike rentals slightly, while the time and temp groups reduce it largely.
• Figure 5: Schematic overview over casual_ordering = list($\tau_1$ = 1, $\tau_2$ = c(2,3), $\tau_3$ = c(4,5,6,7)) without and with confounding specified in the second component.