Pairwise graph gains track length and local signs, not topology
Original title: Interpreting Supervised Pairwise Graph Representations: Local Evidence, Confounding, and Provable Post-hoc Correction
After controlling for sequence length, the reported edge over model confidence disappears; topology shifts ROC-AUC by at most 0.004.
The paper argues that accuracy of supervised pairwise graphs does not show that higher-order topology carries the signal. In a controlled Transformer error-detection task, those graphs beat attention-as-adjacency (ROC-AUC 0.839 versus 0.559; a no-confidence pair is 0.687 versus 0.559), but the advantage over model confidence vanishes after controlling for sequence length, recoverable from the pooled hidden state at R^2=0.9995. On published HCP Emotion ensemble graphs, edge magnitude is near chance and topology-aware changes move ROC-AUC by at most 0.004, while a validation-tuned three-layer GCN does not beat a signed node-summary baseline. A Fisher post-hoc corrector meets its stated guarantees; a separate cross-attention analysis finds a redundant reversal circuit plus EOS-timing and copy failures.
- Sequence length is almost perfectly recoverable from the pooled hidden state (R^2=0.9995).
- Only sign-separated degree shows small positive changes across all tested graph sizes.
- In a near-error-free fMRI calibration regime the Fisher bound becomes statistically uninformative.
- Signed node aggregation already acts as an ensemble of supervised pairwise predictions.
From the original article
1 min read at arxiv.orgSupervised pairwise graph representations convert pairwise predictive relations into sample-specific graphs, but graph-level accuracy does not by itself show that higher-order topology carries the signal. In a controlled Transformer error-detection task, supervised pairwise graphs outperform attention-as-adjacency (ROC-AUC $0. 839$ versus $0.
By Ilya Brekin · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.