A new paper on arXiv introduces Graph Additive Models, a family of architectures built for interpretable graph learning. The core proposal, Graph Neural Additive Networks (GNANs), is designed to make predictions transparent by construction rather than relying on post-hoc explanations.

The abstract describes GNANs as combining per-feature shape functions with simple distance-based terms. This structure lets the model separate the influence of individual node features from the effects of graph topology and feature interactions. The title frames this as disentangling feature, topology, and interactions, which is the central contribution.

Because the source is a single arXiv abstract, the details are limited. The authors position GNANs as an effective alternative to opaque graph neural networks, but the full methodology and experimental results are not yet available in the abstract itself.