Parameter efficient interpretable neural additive model for graph data

We generate global text-based explanations using representative nodes (exemplars) in the embedding space. The exemplars are selected via coverage maximization, and their signatures are explained using natural language rules from a self-refining LLM.