Source-linked AI summary
Model Agnostic Supervised Local Explanations
Gregory Plumb, Denali Molitor, Ameet Talwalkar
TL;DR
Interpretability systems must address example-based, local, and global explanations for increasingly opaque models used in critical decisions. MAPLE combines local linear modeling with random-forest neighborhoods and feature selection, achieving accurate prediction, faithful explanations, and global-pattern detection. Its local training distribution also helps diagnose when local explanations are unreliable.
Problem
Opaque models used in critical decisions require explanation systems that capture example-based, local, and global behavior.
Method
MAPLE combines local linear models with random forests as a supervised neighborhood approach and feature-selection method.
Results
MAPLE is at least as accurate as several predictive baselines, provides faithful self-explanations, outperforms LIME for black-box response prediction, and detects global patterns.
Takeaways & Limitations
MAPLE provides example-based and local explanations while using its local training distribution to diagnose limitations in those explanations.
Takeaways & Limitations
A zero local coefficient can mean either local inactivity or a significant feature whose effect is global rather than local.
Abstract
from arXiv · showhide
Model interpretability is an increasingly important component of practical machine learning. Some of the most common forms of interpretability systems are example-based, local, and global explanations. One of the main challenges in interpretability is designing explanation systems that can capture aspects of each of these explanation types, in order to develop a more thorough understanding of the model. We address this challenge in a novel model called MAPLE that uses local linear modeling techniques along with a dual interpretation of random forests (both as a supervised neighborhood approach and as a feature selection method). MAPLE has two fundamental advantages over existing interpretability systems. First, while it is effective as a black-box explanation system, MAPLE itself is a highly accurate predictive model that provides faithful self explanations, and thus sidesteps the typical accuracy-interpretability trade-off. Specifically, we demonstrate, on several UCI datasets, that MAPLE is at least as accurate as random forests and that it produces more faithful local explanations than LIME, a popular interpretability system. Second, MAPLE provides both example-based and local explanations and can detect global patterns, which allows it to diagnose limitations in its local explanations.
1 Introduction
MAPLE combines example-based, local, and global perspectives to explain models while retaining predictive accuracy. Its local training distributions also help detect global patterns and judge when exemplar explanations apply.
- MAPLE combines local linear models with random forests’ supervised-neighborhood and feature-selection roles.
- MAPLE provides example-based and local explanations while remaining a highly accurate predictive model.
- MAPLE’s local training distributions detect global patterns that local explanations may miss or model poorly.
- MAPLE uses local training distributions to select appropriate exemplar explanations for new test points.
- MAPLE can serve as a black-box explainer and produces more faithful explanations than LIME.
2 Background and Related Work
The background defines example-based, local, and global explanations and identifies challenges in evaluating and transferring local explanations. It motivates MAPLE’s use of random forests for supervised neighborhoods and feature importance.
- Interpretability is defined as human-simulatable calculation with semantically meaningful features.
- A local explanation is an interpretable function that approximates a predictive model within a neighborhood around an input.
- MAPLE’s related-work position follows LIME’s local sparse-linear modeling while addressing explanation fidelity and transfer concerns.
- Local explanation evaluation should assess how well explanations predict model responses under input perturbations, not only at the original point.
- Global explanations use rules covering all or defined regions of the input space, trading coverage against precision.
- Random forests support MAPLE as both a supervised neighborhood-selection method and a source of global feature importance.
3 MAPLE
MAPLE combines random-forest neighborhood weights with feature selection and local linear modeling to produce supervised local predictions and explanations.
- MAPLE: MAPLE uses random forests both to define local training distributions and to score features for selection.Training-point weights are based on shared leaf membership across trees; DStump scores root-split features by impurity reduction.
- MAPLE: SILO estimates local predictions with weighted linear regression over training points weighted by random-forest neighborhood proximity.The weights form a diagonal matrix used in the local weighted regression problem.
- MAPLE: The component methods have separate theoretical guarantees under stated assumptions, but the paper emphasizes interpretability applications rather than statistical properties.SILO consistency requires certain regularity conditions, while DStump feature identification assumes a general additive model.
- MAPLE: MAPLE fits a weighted local linear model using the selected features, together with a constant term for the local bias.The selected feature count d is chosen by evaluating validation accuracy across d = 1, . . . , p.
- MAPLE: MAPLE can also use gradient boosted regression trees to generate local training distributions and feature scores.These quantities are then provided to MAPLE for local modeling.
4 MAPLE as an Explanation System
MAPLE generates local explanations from weighted training distributions and uses those distributions to detect global patterns and select applicable exemplar explanations.
- Generating Explanations and Detecting Global Patterns: MAPLE can explain either its own predictions or a black-box model’s predictions using a local linear model and local training distribution.The distinction is whether MAPLE is fit to the response variable or to the predictive model’s predicted response.
- Generating Explanations and Detecting Global Patterns: Figure 2 shows influential-point distributions that are smooth and centered for linear effects but wider, disjoint, or non-centered near global patterns.The SIL and Step datasets display distribution changes associated with flatter regions and discontinuities.
- Generating Explanations and Detecting Global Patterns: A zero local coefficient may indicate either local inactivity or a globally patterned feature that the local explanation does not capture.MAPLE therefore diagnoses whether zero coefficients conceal globally significant features.
- Generating Explanations and Detecting Global Patterns: MAPLE uses boxplots and feature-wise grid searches over its local training distribution to diagnose possible global patterns.Skewed distributions suggest proximity to a global pattern; otherwise, grid-search behavior is examined across the feature range.
- Picking an Exemplar Explanation: For exemplar explanations, MAPLE evaluates how likely a proposed test point is under each exemplar’s local training distribution.This provides a principled basis for choosing among exemplar explanations.
- Picking an Exemplar Explanation: If exemplar distributions do not cover the input space, MAPLE can identify uncovered test points and avoid applying an exemplar.It can also detect when multiple exemplar explanations are equally applicable.
5 Experimental Results
Experiments on UCI datasets evaluate MAPLE’s predictive accuracy, self-explanations, black-box explanations, and use of influential training points to reveal global patterns. MAPLE generally matches or exceeds tree-ensemble baselines and outperforms LIME on most tested settings, with performance depending on neighborhood scale and dimensionality.
- 5.1 Accuracy on UCI datasets: MAPLE was generally at least as accurate as random forests, GBRT, and SILO, often performing better; Music was the sole exception.Experiments used 50/25/25 training, validation, and testing splits on several UCI datasets, with RMSE reported for predictive performance.
- 5.2 Faithful Self-Explanations: MAPLE’s local linear models provided good self-explanations under the causal metric when compared with LIME for σ = 0.1.The causal metric used perturbed points sampled from N(x, σI), with five perturbations per test point and squared l2 loss.
- 5.3 MAPLE as a Black-box Explainer: MAPLE produced more accurate local explanations than LIME in the black-box setting for all but the Crimes dataset, where the difference was not statistically significant.The black-box experiments fit explanation systems to an SVR model’s predicted responses and evaluated RMSE using the causal metric.
- 5.3 MAPLE as a Black-box Explainer: With σ = 0.25, MAPLE significantly outperformed LIME on Autompgs, Happiness, Housing, and Winequality-red, while LIME significantly outperformed MAPLE on Communities, Crimes, and Music.The larger neighborhood was considered unreasonably large for local explanations in high-dimensional problems, where the neighborhood expands further.
- 5.4 Using Influential Training Points: Influential training-point distributions changed smoothly for continuous effects, became concentrated near steep transitions, and shifted abruptly around discontinuities.The experiments used n = 200 samples in [0, 1]5, fitted MAPLE to random-forest predictions, and plotted the 20 most influential points across an active-feature grid.
6 Conclusion and Future Work
The conclusion presents MAPLE as effective both as a predictive model and an explanation system, while identifying ways to extend its explanatory and modeling capabilities.
- MAPLE is effective as both a predictive model and an explanation system.
- MAPLE’s local training distribution addresses detecting and modeling global patterns in local explanations.Global patterns are effects that local explanations may fail to detect or model, including discontinuities and strong effects confined to small regions.
- MAPLE’s local training distribution helps determine whether an exemplar explanation applies to a new test point.
- Future work includes using influence-function methods with MAPLE to improve accuracy or identify interesting data points.The paper specifically mentions leverage and Cook’s distance as possible measures.
- Future work includes local feature selection through tree-path impurity reductions and alternative methods for defining similarity weights.