Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Viewing the process of generating counterfactuals as a source of knowledge: a new approach for explaining classifiers

Vincent Lemaire, Nathan Le Boudec, Victor Guyomard, Françoise Fessant

TL;DR

This paper tackles explainability by reframing counterfactual generation as a knowledge-creating process. It builds a knowledge base for the Naive Bayes classifier by encoding the impact of feature changes with $\Delta$ values and demonstrates that changes are additive, enabling tractable trajectory construction toward a target class. The Telco churn case shows how clustering Delta-based profiles yields actionable, trajectory-driven preventive and reactive interventions. Overall, the work provides a practical framework to convert local counterfactual explanations into global, reusable knowledge for explainability, with potential extensions to domains beyond churn.

Abstract

There are now many explainable AI methods for understanding the decisions of a machine learning model. Among these are those based on counterfactual reasoning, which involve simulating features changes and observing the impact on the prediction. This article proposes to view this simulation process as a source of creating a certain amount of knowledge that can be stored to be used, later, in different ways. This process is illustrated in the additive model and, more specifically, in the case of the naive Bayes classifier, whose interesting properties for this purpose are shown.

Viewing the process of generating counterfactuals as a source of knowledge: a new approach for explaining classifiers

TL;DR

This paper tackles explainability by reframing counterfactual generation as a knowledge-creating process. It builds a knowledge base for the Naive Bayes classifier by encoding the impact of feature changes with values and demonstrates that changes are additive, enabling tractable trajectory construction toward a target class. The Telco churn case shows how clustering Delta-based profiles yields actionable, trajectory-driven preventive and reactive interventions. Overall, the work provides a practical framework to convert local counterfactual explanations into global, reusable knowledge for explainability, with potential extensions to domains beyond churn.

Abstract

There are now many explainable AI methods for understanding the decisions of a machine learning model. Among these are those based on counterfactual reasoning, which involve simulating features changes and observing the impact on the prediction. This article proposes to view this simulation process as a source of creating a certain amount of knowledge that can be stored to be used, later, in different ways. This process is illustrated in the additive model and, more specifically, in the case of the naive Bayes classifier, whose interesting properties for this purpose are shown.
Paper Structure (24 sections, 1 theorem, 12 equations, 4 figures, 2 tables)

This paper contains 24 sections, 1 theorem, 12 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Concepts
  3. Countrefactual and semi-factual example
  4. Informativeness and actionability of variables
  5. The concept of trajectories
  6. Optimised search for counterfactuals in the case of the naive Bayes classifier
  7. Reminders on the naive Bayes classifier
  8. Criteria to be optimised in the search for counterfactuals
  9. Creation and use of a knowledge base
  10. Creation a knowledge base
  11. Generation of counterfactuals with certain properties
  12. Minimising the number of changes -
  13. Taking account of business constraints or criteria -
  14. Additional usable knowledge
  15. Preventive and reactive actions -
  16. ...and 9 more sections

Key Result

proposition thmcounterproposition

If we take $X$ and $X'$ as two elements of the input space $\mathcal{X}$, we show that for a two-class classification problem, searching for counterfactuals of $X$ amounts to examining the evolution of the value of $\Delta$ when we change some of the values of $X$ to $X'$, such that:

Figures (4)

  • Figure 1: Illustration of a counterfactual and a semi-factual. The red dots represent initial examples ($X$). The orange represents a semi-factual, the purple dot represents a counterfactual and the white line represents the decision boundary between the red and green classes.
  • Figure 2: Illustration of a counterfactual and a semi-factual. The red dots represent initial examples ($X$). The orange dot represents a semi-factual, the purple dot represents a counterfactual and the white line represents the decision boundary between the red and green classes.
  • Figure 3: Illustration of two counterfactuals: one achieved in 1 step, the second in 3 steps
  • Figure 4: Average profile of individuals in clusters represented as histograms. The names of the values on the abscissa refer to the number of the variables and the number of the intervals (groups) described above. For example, '3I2' refers to the third variable ('Contract') and its second interval / group of values ('Twoyear'). The ordinate values are the mean values of the cluster ($\Delta$).

Theorems & Definitions (2)

  • definition thmcounterdefinition
  • proposition thmcounterproposition