CArLA

Scalability

Arg&App 2025AAAI 2026 ORAL

Heterogeneous Graph Neural Networks for Assumption-Based Argumentation

Uses heterogeneous graph neural networks to approximate credulous acceptance and reconstruct stable extensions for large ABA frameworks.

The problem

Exact computation of stable extensions in assumption-based argumentation is intractable for large frameworks, limiting interactive use.

The approach

ABA frameworks are encoded as dependency graphs with heterogeneous node and edge types. Residual convolution and attention architectures learn node representations and predict credulous acceptance.

What the work establishes

  • ABAGCN and ABAGAT outperform an adapted graph-neural baseline, reaching node-level F1 up to 0.71 on ICCMA instances.
  • A sound polynomial-time reconstruction algorithm achieves extension F1 above 0.85 on small frameworks.
  • The reconstruction approach maintains approximately 0.58 F1 on large frameworks.

Why it matters to CArLA

The work addresses the scalability objective by providing learned approximations while retaining a route back to valid argumentative extensions.

Contributor acknowledgement

  • Preesha Gehlot — author and supervised student contributor

Citation

Gehlot, P., Rapberger, A., Russo, F., & Toni, F. (2026). Heterogeneous Graph Neural Networks for Assumption-Based Argumentation. Proceedings of AAAI, 40(23), 19117–19125. https://doi.org/10.1609/aaai.v40i23.38985