[Paper Review] REPID: Regional Effect Plots with implicit Interaction Detection

Julia Herbinger, Bernd Bischl+1 more|arXiv (Cornell University)|Feb 15, 2022

Explainable Artificial Intelligence (XAI)4 citations

TL;DR

REPID introduces a model-agnostic framework for regional effect plots that detect and quantify feature interactions implicitly by partitioning the feature space into interpretable subgroups via decision trees. It reduces aggregation bias in global explanations by isolating feature effects within regions, demonstrating superior interaction detection and visualization on simulation and real-world datasets like California housing and diabetes data.

ABSTRACT

Machine learning models can automatically learn complex relationships, such as non-linear and interaction effects. Interpretable machine learning methods such as partial dependence plots visualize marginal feature effects but may lead to misleading interpretations when feature interactions are present. Hence, employing additional methods that can detect and measure the strength of interactions is paramount to better understand the inner workings of machine learning models. We demonstrate several drawbacks of existing global interaction detection approaches, characterize them theoretically, and evaluate them empirically. Furthermore, we introduce regional effect plots with implicit interaction detection, a novel framework to detect interactions between a feature of interest and other features. The framework also quantifies the strength of interactions and provides interpretable and distinct regions in which feature effects can be interpreted more reliably, as they are less confounded by interactions. We prove the theoretical eligibility of our method and show its applicability on various simulation and real-world examples.

Motivation & Objective

• To address aggregation bias in global interpretation methods like partial dependence plots, which can misrepresent feature effects when interactions are present.
• To develop a method that identifies interpretable subgroups in the feature space where feature effects are less confounded by interactions.
• To quantify interaction strength between a feature of interest and other features, enabling ranking of interactions by importance.
• To combine local interpretability (via subgroups) with global insights, offering a balanced approach between local and global explanation methods.
• To provide a theoretically grounded, reproducible framework for detecting and visualizing interactions in complex machine learning models.

Proposed method

• The method uses a decision tree to partition the feature space into distinct, interpretable subgroups based on the feature of interest and its potential interactions.
• Within each subgroup, regional effect plots (REPs) are computed to visualize the marginal effect of the feature of interest, minimizing confounding from interactions.
• A novel interaction detection measure is proposed to quantify the strength of interactions between the feature of interest and other features, based on variance decomposition in predictions across subgroups.
• The method is model-agnostic and can be applied to any black-box model after training.
• A global interaction index is derived from SHAP interaction values for comparative evaluation.
• The framework integrates ICE curve visualization to detect heterogeneous effects, which are then localized and quantified within subgroups.

Experimental results

Research questions

• RQ1 How can we detect and quantify feature interactions in machine learning models without relying on global aggregation that obscures heterogeneous effects?
• RQ2 Can we identify interpretable subgroups in the feature space where the effect of a feature of interest is more reliably interpretable and less confounded by interactions?
• RQ3 How does the proposed interaction detection measure compare to established indices like the H-statistic and Greenwell’s interaction index in terms of accuracy and interpretability?
• RQ4 To what extent can regional effect plots improve the reliability of global model explanations in the presence of complex interactions?
• RQ5 Can the method uncover hidden biases or subgroup-specific effects in real-world datasets, such as in housing or healthcare applications?

Key findings

• REPID successfully identifies subgroups where the effect of a feature of interest varies significantly, reducing aggregation bias in global explanations.
• For the California housing dataset, the interaction importance between Longitude and Latitude was 0.49, and with Ocean proximity was 0.18, with an R² of interaction strength at 0.67.
• In the diabetes dataset, Skin fold thickness showed a non-monotonic effect on diabetes risk, with distinct patterns in subgroups defined by Glu (interaction importance 0.29) and Age (0.09), with R²_int = 0.49.
• The method outperformed the H-statistic and Greenwell’s interaction index in detecting and ranking interactions, especially in cases with non-linear and heterogeneous effects.
• The framework revealed that the effect of Longitude on house value decreased with increasing latitude and varied by proximity to the ocean, which was not apparent in global PD plots.
• The implementation is publicly available, ensuring full reproducibility and broad accessibility for researchers and practitioners.