Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Heterogeneous Graph Neural Networks for Assumption-Based Argumentation

Preesha Gehlot, Anna Rapberger, Fabrizio Russo, Francesca Toni

TL;DR

This work tackles the computational intractability of credulous acceptance under stable semantics in Assumption-Based Argumentation (ABA) by introducing native heterogeneous Graph Neural Networks (GNNs) operating on a faithful dependency-graph encoding of ABA frameworks. It presents two architectures, ABAGCN and ABAGAT, which leverage relation-specific GCN and GAT layers to learn node embeddings for assumptions, rules, and claims, achieving strong accuracy on ICCMA 2023 benchmarks and synthetic ABAFs. A sound polynomial-time extension-reconstruction algorithm uses the predictor to assemble stable extensions, delivering high F1 on small instances and substantial runtime gains over exact solvers on large frameworks. The approach demonstrates that GNN-based approximate reasoning can significantly bridge the gap between scalability and accuracy in structured argumentation, with practical implications for learning ABAFs and causal/discovery tasks.

Abstract

Assumption-Based Argumentation (ABA) is a powerful structured argumentation formalism, but exact computation of extensions under stable semantics is intractable for large frameworks. We present the first Graph Neural Network (GNN) approach to approximate credulous acceptance in ABA. To leverage GNNs, we model ABA frameworks via a dependency graph representation encoding assumptions, claims and rules as nodes, with heterogeneous edge labels distinguishing support, derive and attack relations. We propose two GNN architectures - ABAGCN and ABAGAT - that stack residual heterogeneous convolution or attention layers, respectively, to learn node embeddings. Our models are trained on the ICCMA 2023 benchmark, augmented with synthetic ABAFs, with hyperparameters optimised via Bayesian search. Empirically, both ABAGCN and ABAGAT outperform a state-of-the-art GNN baseline that we adapt from the abstract argumentation literature, achieving a node-level F1 score of up to 0.71 on the ICCMA instances. Finally, we develop a sound polynomial time extension-reconstruction algorithm driven by our predictor: it reconstructs stable extensions with F1 above 0.85 on small ABAFs and maintains an F1 of about 0.58 on large frameworks. Our work opens new avenues for scalable approximate reasoning in structured argumentation.

Heterogeneous Graph Neural Networks for Assumption-Based Argumentation

TL;DR

This work tackles the computational intractability of credulous acceptance under stable semantics in Assumption-Based Argumentation (ABA) by introducing native heterogeneous Graph Neural Networks (GNNs) operating on a faithful dependency-graph encoding of ABA frameworks. It presents two architectures, ABAGCN and ABAGAT, which leverage relation-specific GCN and GAT layers to learn node embeddings for assumptions, rules, and claims, achieving strong accuracy on ICCMA 2023 benchmarks and synthetic ABAFs. A sound polynomial-time extension-reconstruction algorithm uses the predictor to assemble stable extensions, delivering high F1 on small instances and substantial runtime gains over exact solvers on large frameworks. The approach demonstrates that GNN-based approximate reasoning can significantly bridge the gap between scalability and accuracy in structured argumentation, with practical implications for learning ABAFs and causal/discovery tasks.

Abstract

Assumption-Based Argumentation (ABA) is a powerful structured argumentation formalism, but exact computation of extensions under stable semantics is intractable for large frameworks. We present the first Graph Neural Network (GNN) approach to approximate credulous acceptance in ABA. To leverage GNNs, we model ABA frameworks via a dependency graph representation encoding assumptions, claims and rules as nodes, with heterogeneous edge labels distinguishing support, derive and attack relations. We propose two GNN architectures - ABAGCN and ABAGAT - that stack residual heterogeneous convolution or attention layers, respectively, to learn node embeddings. Our models are trained on the ICCMA 2023 benchmark, augmented with synthetic ABAFs, with hyperparameters optimised via Bayesian search. Empirically, both ABAGCN and ABAGAT outperform a state-of-the-art GNN baseline that we adapt from the abstract argumentation literature, achieving a node-level F1 score of up to 0.71 on the ICCMA instances. Finally, we develop a sound polynomial time extension-reconstruction algorithm driven by our predictor: it reconstructs stable extensions with F1 above 0.85 on small ABAFs and maintains an F1 of about 0.58 on large frameworks. Our work opens new avenues for scalable approximate reasoning in structured argumentation.

Paper Structure

This paper contains 26 sections, 5 theorems, 13 equations, 3 figures, 3 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Preliminaries
  4. Computational Argumentation
  5. Assumption-based Argumentation
  6. Graph Neural Networks
  7. GCNs
  8. GATs
  9. Predicting Acceptance with GNNs
  10. Dependency Graph
  11. Neural Architecture
  12. Extension Reconstruction
  13. Empirical Evaluation
  14. Data
  15. Baseline
  16. ...and 11 more sections

Key Result

Proposition 3.6

Let $\mathcal{D}\!=\!(\mathcal{L},\mathcal{R},\mathcal{A},\overline{\space})$ be an ABAF and $F_{\mathcal{D}}$ its associated AF. If $E\!\in\! {\mathit{stb}}(F_{\mathcal{D}})$ then $\bigcup_{S\vdash p\!\in E}S\!\in\! {\mathit{stb}}(\mathcal{D})$; if $S\!\in\! {\mathit{stb}}(\mathcal{D})$ then $\{S'\

Figures (3)

  • Figure 1: Diagram of the model architecture, including both the feature extraction (using the dependency graph described in \ref{['sec:dep_g']}) and the GNN described in \ref{['sec:neural_arch']}. The Heterogeneous Graph Layer module would apply Convolutional Layers for ABAGCN, or Attention Layers for ABAGAT.
  • Figure 2: Model comparison according to F1, Precision, Recall and Accuracy on different cuts of data: small ICCMA ABAFs (with less than 100 elements) in panel (a) to be able to compare to the AFGCN baseline; Comparison between the full ICCMA test set and the test set augmented with our generated data in panel (b).
  • Figure 3: F1 score of the extension by ABAF size.

Theorems & Definitions (17)

  • Definition 3.1
  • Definition 3.2
  • Definition 3.3
  • Definition 3.4
  • Example 3.5
  • Definition 3.6
  • Proposition 3.6
  • Definition 4.1
  • Example 4.3
  • Proposition 5.0
  • ...and 7 more