Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

T-Explainer: A Model-Agnostic Explainability Framework Based on Gradients

Evandro S. Ortigossa, Fábio F. Dias, Brian Barr, Claudio T. Silva, Luis Gustavo Nonato

TL;DR

This paper addresses the opacity of black-box ML models by introducing T-Explainer, a deterministic, model-agnostic, additive explainer grounded in Taylor expansion. It estimates local feature attributions via centered finite differences and a dedicated $h$-optimizer, yielding a surrogate $g_{ extbf{x}}(\textbf{z}')$ with $\phi_0 = E[f(\textbf{X})]$ and $\phi_i = \partial f(\textbf{x})/\partial x_i$, while aiming to satisfy Local Accuracy, Missingness, and Consistency. The authors formalize new stability and faithfulness metrics (RIS, ROS, RES, PGI) and validate T-Explainer against SHAP, LIME, and gradient-based methods on synthetic and real-world datasets, including non-differentiable models. Results show strong stability and competitive faithfulness across tasks, with a practical Python framework that integrates T-Explainer with other XAI tools. The work provides a scalable, robust XAI alternative and outlines plans to extend to regression, multiclass, and more complex data modalities.

Abstract

The development of machine learning applications has increased significantly in recent years, motivated by the remarkable ability of learning-powered systems to discover and generalize intricate patterns hidden in massive datasets. Modern learning models, while powerful, often exhibit a complexity level that renders them opaque black boxes, lacking transparency and hindering our understanding of their decision-making processes. Opacity challenges the practical application of machine learning, especially in critical domains requiring informed decisions. Explainable Artificial Intelligence (XAI) addresses that challenge, unraveling the complexity of black boxes by providing explanations. Feature attribution/importance XAI stands out for its ability to delineate the significance of input features in predictions. However, most attribution methods have limitations, such as instability, when divergent explanations result from similar or the same instance. This work introduces T-Explainer, a novel additive attribution explainer based on the Taylor expansion that offers desirable properties such as local accuracy and consistency. We demonstrate T-Explainer's effectiveness and stability over multiple runs in quantitative benchmark experiments against well-known attribution methods. Additionally, we provide several tools to evaluate and visualize explanations, turning T-Explainer into a comprehensive XAI framework.

T-Explainer: A Model-Agnostic Explainability Framework Based on Gradients

TL;DR

This paper addresses the opacity of black-box ML models by introducing T-Explainer, a deterministic, model-agnostic, additive explainer grounded in Taylor expansion. It estimates local feature attributions via centered finite differences and a dedicated -optimizer, yielding a surrogate with and , while aiming to satisfy Local Accuracy, Missingness, and Consistency. The authors formalize new stability and faithfulness metrics (RIS, ROS, RES, PGI) and validate T-Explainer against SHAP, LIME, and gradient-based methods on synthetic and real-world datasets, including non-differentiable models. Results show strong stability and competitive faithfulness across tasks, with a practical Python framework that integrates T-Explainer with other XAI tools. The work provides a scalable, robust XAI alternative and outlines plans to extend to regression, multiclass, and more complex data modalities.

Abstract

The development of machine learning applications has increased significantly in recent years, motivated by the remarkable ability of learning-powered systems to discover and generalize intricate patterns hidden in massive datasets. Modern learning models, while powerful, often exhibit a complexity level that renders them opaque black boxes, lacking transparency and hindering our understanding of their decision-making processes. Opacity challenges the practical application of machine learning, especially in critical domains requiring informed decisions. Explainable Artificial Intelligence (XAI) addresses that challenge, unraveling the complexity of black boxes by providing explanations. Feature attribution/importance XAI stands out for its ability to delineate the significance of input features in predictions. However, most attribution methods have limitations, such as instability, when divergent explanations result from similar or the same instance. This work introduces T-Explainer, a novel additive attribution explainer based on the Taylor expansion that offers desirable properties such as local accuracy and consistency. We demonstrate T-Explainer's effectiveness and stability over multiple runs in quantitative benchmark experiments against well-known attribution methods. Additionally, we provide several tools to evaluate and visualize explanations, turning T-Explainer into a comprehensive XAI framework.
Paper Structure (18 sections, 14 equations, 3 figures, 19 tables, 1 algorithm)

This paper contains 18 sections, 14 equations, 3 figures, 19 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related work
  3. The T-Explainer
  4. T-Explainer Properties
  5. Local Accuracy
  6. Missingness
  7. Consistency
  8. T-Explainer: Computational Aspects
  9. Handling Categorical Data
  10. Quantitative Metrics
  11. Experiments
  12. Experimental Setup
  13. Synthetic Data
  14. Real-world Data
  15. Computational Performance
  16. ...and 3 more sections

Figures (3)

  • Figure 1: T-Explainer pipeline.
  • Figure 2: Behavior of XAI methods explaining samples predicted with the 3H-NN model trained on the 20-FT dataset when running RIS and ROS metrics multiple times. The y-axes are on a logarithmic scale for better visualization.
  • Figure 3: Running times (in log scale) of T-Explainer and SHAP explainers applied to the 3H-NN model trained on 16-dimensional synthetic data.