Arguing and Explaining Classifications

Leila Amgoud, Mathieu SerrurierArgMAS 2007 · DOI

Abstract

Traditional machine learning models often operate as "black boxes," providing classifications without transparent justifications. This paper, originally presented at the 2007 Workshop on Argumentation in Multi-Agent Systems (ArgMAS), introduces a formal argumentation-based framework that treats classification as a structured debate. By representing hypotheses and training data as arguments that can attack one another, the system provides both a final label and a human-readable explanation. This approach gracefully handles inconsistent data and generalizes classical Version Space learning.

Introduction

Standard classification models, such as decision trees or neural networks, typically select a single "best" hypothesis to label new data. This approach suffers from two primary limitations: a lack of transparency regarding the decision-making process and difficulty handling inconsistent or noisy training data. When training data contains contradictory examples, many classical algorithms struggle to provide a robust justification for their choices.

To address these gaps, authors Leila Amgoud and Mathieu Serrurier apply formal argumentation theory to the classification task. Instead of selecting a single winning hypothesis, the framework allows all possible hypotheses to "argue" for different classifications. The final decision is reached by evaluating which arguments are "acceptable" under logical scrutiny.

Methodology

The proposed system is built upon Dung’s Abstract Argumentation Framework, a cornerstone of symbolic AI. Classification is translated into a debate through several key components:

  • Arguments: An argument is defined as a triplet consisting of a hypothesis, an example, and a class. Arguments can be derived directly from training examples (considered strong evidence) or from general hypotheses (considered weaker).
  • Attacks: The framework defines two types of conflict:
  • Rebutting: Occurs when two arguments assign different classes to the same example.
  • Undercutting: Occurs when a training example proves a hypothesis is incorrect (e.g., the hypothesis predicts "Sunny" but the training data records "Not Sunny" for the same conditions).
  • Acceptability Semantics: The framework utilizes mathematical "extensions"—sets of arguments that are internally consistent and can defend themselves against attacks—to determine justified classifications.
  • Decision Rules: Two primary methods for selecting a class are proposed: a Universal Vote, where a class is chosen only if supported by arguments in every winning set, and a Majority Vote, where the class supported by the majority of winning sets is selected.

Case Study: The "Sunny Day" Example

As a theoretical exploration of formal logic, the paper utilizes a synthetic "Sunny Day" concept learning example to demonstrate the framework's mechanics. This toy dataset allows for a step-by-step walkthrough of how arguments are formed, how they conflict, and how the grounded extension (the set of undisputed arguments) is calculated.

The "Sunny Day" synthetic dataset used to demonstrate the argumentation framework.

By using features like pressure, temperature, and humidity, the authors show how the framework navigates the logic of classification even when data might be sparse or inconsistent.

Key Insights and Results

The argumentation framework offers several significant advantages over traditional methods:

  1. Classification as Explanation: Every classification is backed by a specific argument, providing a built-in explanation. A decision is not merely a label but a statement: "Class X is chosen because Hypothesis A supports it, and Hypothesis A is consistent with Training Examples Y and Z."
  2. Robustness to Inconsistency: Unlike Mitchell’s classical Version Space theory, which can fail if training data is inconsistent, this framework remains functional by weighing conflicting evidence to find the most justifiable answer.
  3. Theoretical Unification: A key result of the paper is the proof that this framework is a generalization of Version Space theory. Proposition 11 demonstrates that the "grounded extension" can be used to characterize the upper and lower bounds of the version space, effectively unifying two different fields of AI.

Conclusion and Future Work

While the framework provides a rigorous logical foundation for explainable AI, challenges remain regarding computational complexity. Finding preferred extensions in an argumentation graph is NP-hard, necessitating optimization for large-scale applications. Future work includes testing the framework on real-world datasets and developing "proof theories" to efficiently check if a specific classification is justified without exhaustive calculation of all arguments.

Summary updated

BibTeX
@inproceedings{amgoud2007arguing,
  author = {Leila Amgoud and Mathieu Serrurier},
  title = {{Arguing and Explaining Classifications}},
  booktitle = {ArgMAS},
  year = {2007},
  doi = {10.1007/978-3-540-78915-4\_11},
}
Figure