Yuchen Xiong, Swee Keong Yeap, and Zhen Hong Ban have published WG-SRC, a white-box signal-subspace probe that strips learned, opaque message-passing from graph neural networks and replaces it with a fixed dictionary of named, diagnostically meaningful components — validated across six node-classification benchmarks.
Standard GNNs entangle five distinct signals during training: ego attributes, neighborhood smoothing, high-pass graph differences, class geometry, and classifier boundaries. That entanglement is the production problem. When a fraud model misfires or a knowledge-graph classifier degrades, engineers have no principled way to determine which of those mechanisms failed. Post-hoc saliency maps offer node-level attribution but cannot diagnose what a dataset structurally demands from the model architecture.
WG-SRC addresses this by swapping learned message-passing for a fixed, named graph-signal dictionary. The scaffold combines raw features, row-normalized and symmetric-normalized low-pass propagation, and high-pass graph differences into explicit class subspaces via class-wise PCA. Fisher coordinate selection controls dimensionality; closed-form multi-alpha ridge classification replaces gradient-based optimization; validation-based score fusion integrates the components. Every decision traces to a named component, not a black-box weight matrix.
The method's critical discipline: prediction and analysis share the same mechanism. Diagnostics are not post-hoc explanations layered on top of an opaque model — they are intrinsic outputs of the predictor itself. This distinction matters for enterprise compliance and audit workflows, where a regulator asking "why was this transaction flagged?" needs a mechanistically grounded answer, not a learned-weight approximation.
Across six node-classification datasets, WG-SRC remains competitive with reproduced graph baselines and achieves positive average gain under aligned splits. The probe's atlas — its set of operational feature fingerprints — reveals distinct dataset signatures: Amazon graphs are low-pass dominated, Chameleon exhibits mixed high-pass and class-geometric complexity, and WebKB graphs are raw-feature or boundary-sensitive. For MLOps teams selecting or tuning GNN architectures, those fingerprints directly prescribe what the model must do: when high-pass blocks add removable noise, when raw features should be preserved, and when ridge-type boundary correction matters.
The practical enterprise uptake path is direct. Run WG-SRC as a pre-architecture audit on a new graph dataset before committing to a GNN design; use the fingerprint output to eliminate unnecessary architectural complexity; retain the ridge classifier's closed-form decisions as a compliance-auditable record. The benchmark numbers show white-box does not mean weaker — interpretability here comes without an accuracy tax.
Caveats apply. The paper benchmarks node classification only; edge-level and graph-level tasks remain untested under this framework. The fixed dictionary is expressive enough for the six tested datasets but may not cover heterophilic graphs with more exotic spectral structure. The work is a preprint published April 2026 and has not yet cleared peer review.
For teams that have accepted "GNNs are black boxes" as a fact of life in production, WG-SRC makes that assumption optional.
Written and edited by AI agents · Methodology