Generating Explanations for Conceptual Validation of Graph Neural Networks: An Investigation of Symbolic Predicates...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Graph Neural Networks (GNNs) excel at relational data classification.
• Limited research exists on GNNs' concept learning capabilities and output validation from user and domain perspectives.
• Validatable relational concept learning requires integrating symbolic and statistical machine learning methods.

Purpose of the Study:

• To introduce a benchmark for conceptual validation of GNN classification outputs.
• To develop a framework for generating comprehensible explanations using Inductive Logic Programming (ILP) on GNN relevance outputs.
• To assess the extraction of symbolic concepts from GNN explanations.

Main Methods:

• Utilized symbolic representations of symmetric and non-symmetric figures from the Kandinsky Pattern dataset.
• Developed a novel validation framework combining GNN explainers with ILP.
• Generated human-expected relevance for concepts learned by GNNs.

Main Results:

• Demonstrated the possibility of extracting symbolic concepts from GNN explanations.
• Showcased that extracted concepts are representative of GNN learning.
• Validated the effectiveness of the proposed benchmark and framework.

Conclusions:

• Combining GNNs with ILP facilitates powerful and validatable relational concept learning.
• The developed framework enables the generation of comprehensible and domain-relevant explanations for GNNs.
• Findings open avenues for future research in explainable AI for GNNs.