Learning Gradual Argumentation Frameworks using Meta-heuristics

Nico Potyka, Mohamad Bazo, Jonathan Spieler, Steffen StaabArgML@COMMA 2022 · OpenReview

Title: Learning Gradual Argumentation Frameworks using Meta-heuristics

Introduction

Deep learning models, particularly neural networks, are notoriously difficult to interpret. While they excel at complex tasks, understanding why a model reached a specific conclusion is often obscured by millions of dense parameters. Existing Explainable AI (XAI) methods frequently rely on post-hoc approximations, which may not faithfully represent the model's internal mechanics. This research proposes an alternative: building models using inherently interpretable structures. By exploiting the mathematical equivalence between certain argumentation frameworks and Multi-Layer Perceptrons (MLPs), the authors address the structure learning problem to find the simplest graphs that maintain high predictive accuracy.

The MLP-GAF Paradigm

The researchers focus on Gradual Argumentation Frameworks (GAFs), where nodes represent arguments and weighted edges represent relationships of attack or support. They utilize a specific "MLP-based semantics" where an acyclic GAF corresponds exactly to an MLP with logistic activation functions. In this setup, the "strength" of an argument is calculated in the same manner as a neuron's activation.

Schematic of a GAF and its MLP-based semantics

To facilitate this, continuous data is binarized into logical arguments (e.g., "Petal Width $\in$ [1.3, 2.5]"). This transformation allows the model's decision-making process to be viewed as a transparent debate, where specific features act as arguments that either support or attack a final conclusion. Unlike post-hoc explanation methods, these GAFs are the model, ensuring that the explanation is 100% faithful to the underlying computation.

Theoretical Foundations

A critical aspect of this work is the selection of the MLP-based semantics. The authors demonstrate that this approach satisfies a comprehensive set of desirable properties for argumentation frameworks, such as anonymity, independence, and stability. As shown in the comparison below, the MLP semantics is theoretically robust, meeting nearly all standard criteria for logical consistency, which justifies its use as a foundation for interpretable machine learning.

Comparison of argumentation semantics properties

Optimization via Meta-heuristics

Searching for the optimal sparse graph is an NP-hard problem due to the exponentially large search space. The authors employ two bio-inspired meta-heuristic algorithms to find effective solutions:

  • Genetic Algorithm (GA): Evolves a population of candidate structures through selection, crossover, and mutation.
  • Particle Swarm Optimization (PSO): Moves a "swarm" of candidate solutions through the search space based on local and global best-known positions.

A fitness function guides these algorithms by rewarding high training accuracy while penalizing the number of edges. This pressure toward sparsity ensures the resulting models remain human-comprehensible.

Experimental Results

The approach was evaluated using three classic benchmarks from the UCI Machine Learning Repository: Iris, Mushroom, and Adult Income.

  • Accuracy: On the Iris dataset, the Genetic Algorithm achieved 97.34% accuracy, slightly outperforming standard Decision Trees (95%). Performance on the Mushroom and Adult datasets was comparable to Decision Trees, typically within a 1-2% margin.
  • Complexity: The PSO algorithm was particularly effective at finding extremely sparse models. For the Mushroom dataset, it identified models with an average of only 6.4 edges while maintaining over 97% accuracy.
  • Interpretability: The authors argue that these "flat" GAFs (one hidden layer) are significantly easier for humans to parse than deep decision trees, which can become a confusing "forest" of branches as depth increases.

Conclusion and Future Work

By framing neural network learning as a search for sparse argumentation frameworks, this research provides a path toward "faithful interpretability." The resulting models are not just approximations; they are the computation itself, presented in a human-readable format.

Future work will focus on addressing current limitations, such as moving toward "fuzzy arguments" to better handle continuous data without rigid binarization. Additionally, the authors plan to explore methods for reducing path redundancy in learned graphs and investigating "joint attacks," where combinations of arguments exert unique influences that individual arguments cannot capture alone.

Summary updated

BibTeX
@inproceedings{potyka2022learning,
  author = {Nico Potyka and Mohamad Bazo and Jonathan Spieler and Steffen Staab},
  title = {{Learning Gradual Argumentation Frameworks using Meta-heuristics}},
  booktitle = {ArgML@COMMA},
  year = {2022},
  url = {https://openreview.net/forum?id=vTuDBfwbo0d},
}
Figure