A Dashboard for Simplifying Machine Learning Models using Feature Importances and Spurious Correlation Analysis - Proof-of-Concept Slides

Aug 4, 2026

Upload by Miles

Machine Learning models balance accuracy and explainability, and complex predictors often behave like black boxes in high-stakes settings. This work introduces a dashboard that guides the creation of more interpretable models using Fast-and-Frugal Trees, leveraging feature importance scores and spurious correlation analysis. The workflow supports iterative feature selection and compares tree performance against a complex baseline model, highlighting how spurious correlations can improve feature choice.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages3

Reading Time8 min

AccessFree to download

Introduction

Explainable AI and black-box explainability

Fast-and-Frugal Trees and feature selection

Spurious correlations and misleading feature importance

Approach

Dashboard workflow and data inputs

Model unpickling and prediction pipeline

Feature importance scores and spurious correlation calculation

A Dashboard for Simplifying Machine Learning Models using Feature Importances and Spurious Correlation Analysis T. Cech 1, E. Kohlros2 , W. Scheibel2 and J. Döllner 1 1Digital Engineering Faculty, University of Potsdam, Germany 2Hasso Plattner Institute, Digital Engineering Faculty, University of Potsdam, Germany Abstract Machine Learning models underlie a trade-off between accurracy and explainability. Given a trained, complex model, we contribute a dashboard that supports the process to derive more explainable models, here: Fast-and-Frugal Trees, with further introspection using feature importances and spurious correlation analyses. The dashboard further allows to iterate over the feature selection and assess the trees’performance in comparison to the complex model. CCS Concepts • Human-centered computing → Visualization techniques; • Information systems → Users and interactive retrieval; 1. Introduction In recent years, Artificial Intelligence (AI) models become more prevalent in public discourse. Often, AI models are complex and can be considered a Black Box, as they suffer from a lack of explainability; especially when used in high-stake decision contexts. The field of Explainable AI (XAI) wants to provide techniques that target to explain such Black Box Models, e.g., by analyzing certain properties which are considered decisive for the prediction of the model [LRBB∗ 23] . As Rudin argues, this basic property of many XAI methods lead to some sort of obfuscation since the complex model is not directly explained [Rud19] . She argues that the XAIcommunity should focus on obtaining interpretable models instead. One such model is the Fast-and-Frugal Tree (FFT), a basic Decision Tree with the additional property that there are at max two nodes per level. In the past, Chen et al. have used FFTs in the context of Software Defect Prediction and exemplified, that an FFT model can be competitive to state-of-the-art models [CFKM18] . They argued that for an FFT model to show high quality, one must select few but high-quality features. One way to determine which feature could qualify for such a selection is Feature Importance Scores [LRBB∗ 23] . However, Teng et al. have shown that such feature-based analysis can lead to misleading conclusions when not considering potential Spurious Correlations as exemplified in their VISPUR system [TAL24] . In this work, we present a proof-of-concept for a dashboard that combines Feature Importance Scores with the analysis of Spurious Correlations. For it, we show how spurious correlations could help identify important features for training an FFT on them to obtain a simple yet good enough model. This model is benchmarked side- by-side with the complex model and other FFT variants created by the user. Related Work. Several techniques were proposed for obtaining Feature Importance Scores [LRBB∗ 23] . One example of such a technique is FeatureExplorer by Zhao et al. [ZKM∗ 19] . FeatureExplorer trains consecutive small regression models using selected features for determining an importance score. Additionally, we consider permute-and-predict methods [YSOL09] . For permuteand-predict, the values of several features are permuted, and then consequently the trained model predicts the target based on the permuted feature values. If the model changes its prediction, a feature is considered of high importance since it can not be changed significantly without changing the prediction. Hooker et al. have argued that the permute-and-predict technique can be misleading if the features are highly correlated [HMZ21] . We mitigate this risk by detecting Spurious Correlations especially Simpson’s Paradox [AFL18b] and, additionally, considering a second feature importance score provided by FeatureExplorer. Simpson’s Paradox describes the phenomenon that an overall trend that is present in an aggregated dataset might be missing entirely when disseminating the dataset according to the categories o

FAQs

What problem does the dashboard address in Explainable AI?

It addresses the trade-off between accuracy and explainability by helping users derive simpler, more interpretable models instead of relying on complex black-box predictors.

How does the dashboard use feature importance in building Fast-and-Frugal Trees?

It uses feature importance scores to identify high-quality features, then trains Fast-and-Frugal Trees using the selected top features to obtain a model that is simple yet sufficiently accurate.

Why are spurious correlations important, and how are they detected?

Feature-based analyses can be misleading when spurious correlations exist. The dashboard detects spurious correlations, including Simpson’s Paradox, using a repeated trend-disaggregation approach inspired by Alipourfard et al.

Aug 4, 2026

Upload by Miles