Gradual AA-CBR is a neuro-symbolic AI library implemented in PyTorch that models case-based reasoning as a debate between data points. Each labeled instance acts as an argument advocating for its classification, engaging in structured interactions to determine the outcome of new, unlabeled instances. The framework extends Abstract Argumentation-based Case-Based Reasoning (AA-CBR) with trainable, differentiable mechanisms, allowing the argumentative structure to be learned via gradient descent, similar to how neural networks function.
Install from source. Run the following from the root directory:
pip install .Example usage can be found in the examples/ directory. These examples demonstrate how to construct debates, train models, and make predictions using the Gradual AA-CBR framework.
- 🧠 Neuro-symbolic reasoning: Combines formal argumentation with neural learning.
- 🗣️ Case-based debates: Each data point argues for its label in a structured debate.
- 🔁 Differentiable argumentation semantics: Argumentation structure is learned using backpropagation.
- ⚙️ Modular PyTorch components: Easily integrable into other ML pipelines.
- 📚 Example-driven design: Pre-built examples guide users through usage and training.
Gradual-AA-CBR/
├── src/
│ └── deeparguing/
│ └── ...
├── examples/
│ └── ...
├── data/
│ └── ...
└── README.md
Gradual-AA-CBR is a result of research carried out by the Computational Logic and Argumentation group, at Imperial College London. This repository is based on the following publications:
Adam Gould, Francesca Toni: Neuro-Argumentative Learning with Case-Based Reasoning. arXiv (text, bib)
Kristijonas Cyras, Ken Satoh, Francesca Toni: Abstract Argumentation for Case-Based Reasoning. KR 2016: 549-552 (text, bib)
Oana Cocarascu, Andria Stylianou, Kristijonas Čyras and Francesca Toni: Data-Empowered Argumentation for Dialectically Explainable Predictions. ECAI 2020 (text, bib)
Nico Potyka: Interpreting Neural Networks as Quantitative Argumentation Frameworks. AAAI 2021 (text, bib