Operational Feature Fingerprints of Graph Datasets via a White-Box Signal-Subspace Probe

The key distinction WG-SRC makes is directional: rather than explaining a trained model after the fact, it uses fixed mathematical dictionaries derived from graph signal processing to interrogate what a dataset structurally demands before or during training. That shifts interpretability from a post-hoc audit into something closer to a pre-deployment diagnostic.

This connects directly to the embedding benchmarking work covered here in mid-April ('How Embeddings Shape Graph Neural Networks'), which also used controlled conditions across TU datasets to isolate what architectural choices actually contribute to performance. Both papers are pushing against the same problem: GNN results are hard to attribute because too many variables move at once. WG-SRC takes a more surgical approach by fixing the representation vocabulary entirely. The ORCA interpretability framework for SVMs covered around the same time is a useful parallel, applying a similar philosophy (explicit coordinate decomposition, no retraining required) to a different model class.

The credibility test here is whether WG-SRC's signal components predict performance gaps across the same TU datasets used in the April embedding benchmark. If the two methodologies converge on which datasets are structurally demanding, that would be meaningful cross-validation; if they diverge, at least one set of assumptions needs revisiting.