Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Does the Model Say What the Data Says? A Simple Heuristic for Model Data Alignment

Henry Salgado, Meagan R. Kendall, Martine Ceberio

TL;DR

The paper addresses the interpretability gap by proposing a simple, data-driven baseline to assess model–data alignment using Rubin’s Potential Outcomes-inspired framing. It computes a feature-wise standardized mean difference across outcome groups to rank features by their discriminative power in the data, and then compares this data-derived ranking with model explanations such as SHAP values and decision-tree importances. Experiments on Titanic and Pima Diabetes datasets show moderate to strong agreement between data-driven rankings and model explanations, suggesting that models often learn patterns consistent with the data structure. The approach is computationally efficient and provides practitioners a model-agnostic diagnostic, with limitations including binary classification scope and the need for extensions to multi-class and regression tasks.

Abstract

In this work, we propose a simple and computationally efficient framework for evaluating whether machine learning models align with the structure of the data they learn from; that is, whether the model says what the data says. Unlike existing interpretability methods that focus exclusively on explaining model behavior, our approach establishes a baseline derived directly from the data itself. Drawing inspiration from Rubin's Potential Outcomes Framework, we quantify how strongly each feature separates the two outcome groups in a binary classification task, moving beyond traditional descriptive statistics to estimate each feature's effect on the outcome. By comparing these data-derived feature rankings with model-based explanations, we provide practitioners with an interpretable and model-agnostic method for assessing model-data alignment.

Does the Model Say What the Data Says? A Simple Heuristic for Model Data Alignment

TL;DR

The paper addresses the interpretability gap by proposing a simple, data-driven baseline to assess model–data alignment using Rubin’s Potential Outcomes-inspired framing. It computes a feature-wise standardized mean difference across outcome groups to rank features by their discriminative power in the data, and then compares this data-derived ranking with model explanations such as SHAP values and decision-tree importances. Experiments on Titanic and Pima Diabetes datasets show moderate to strong agreement between data-driven rankings and model explanations, suggesting that models often learn patterns consistent with the data structure. The approach is computationally efficient and provides practitioners a model-agnostic diagnostic, with limitations including binary classification scope and the need for extensions to multi-class and regression tasks.

Abstract

In this work, we propose a simple and computationally efficient framework for evaluating whether machine learning models align with the structure of the data they learn from; that is, whether the model says what the data says. Unlike existing interpretability methods that focus exclusively on explaining model behavior, our approach establishes a baseline derived directly from the data itself. Drawing inspiration from Rubin's Potential Outcomes Framework, we quantify how strongly each feature separates the two outcome groups in a binary classification task, moving beyond traditional descriptive statistics to estimate each feature's effect on the outcome. By comparing these data-derived feature rankings with model-based explanations, we provide practitioners with an interpretable and model-agnostic method for assessing model-data alignment.

Paper Structure

This paper contains 20 sections, 3 equations, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. White-box Models and Post-hoc Techniques
  4. Causal Methods
  5. Data Explainability
  6. Methodology
  7. Evaluation
  8. Datasets
  9. Titanic Dataset
  10. Pima Indians Diabetes Dataset
  11. Data Preprocessing and Experimental Setup
  12. Model Configurations
  13. Neural Network (Multi-Layer Perceptron)
  14. Decision Tree
  15. Alignment Metrics
  16. ...and 5 more sections

Figures (2)

  • Figure 1: Rank comparison scatter plots for the Titanic dataset. Points close to the diagonal indicate agreement between the feature ranking methods.
  • Figure 2: Rank comparison scatter plots for the Diabetes dataset. Both comparisons show strong diagonal clustering, indicating high alignment between SMD-derived feature rankings and model-based importance measures.