REVIEW 3 major objections 3 minor 4 references
Oldie but Goodie: Re-illuminating Label Propagation on Graphs with Partially Observed Features
T0 review · 3 major / 3 minor · reviewed 2026-08-06 · deepseek-v4-flash
Pith's one-line read The paper claims that classical label propagation, paired with feature propagation through a carefully aligned hybrid architecture, is not obsolete: a framework named GOODIE outperforms current state-of-the-art node classifiers both when…
desk verdict The submitted text does not contain the paper: the abstract advertises a graph learning method, but the full text is an unrelated cs.CL paper, so the central empirical claim is unverifiable as supplied. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
Label propagation spreads known labels over graph edges to produce predictions for unlabeled nodes, and feature propagation spreads observed feature values over edges to compensate for missing ones. A GNN-based decoder aligns the hidden embeddings of the label-propagation branch with those of the feature-propagation branch, while a Structure-Feature Attention module learns a weighting between the structural and feature signals. Pseudo-label contrastive learning treats pseudo-labels from the label-propagation branch as positive pairs and differentiates how much each pair contributes. The mechanism that carries the argument is the branch alignment: if the label-propagation branch can emit embeddings in the same space as the feature-propagation branch without losing structural signal, the attention-weighted combination can outperform either branch alone.
What would settle it
Run GOODIE on a standard citation or social graph while masking an increasing fraction of node features, and compare its accuracy with plain label propagation at high masking rates and with a standard GNN at low masking rates; any setting where GOODIE fails to beat both baselines in their respective regimes would contradict the central claim.
Extended reading notes
Core claim
The paper's central discovery, on its own terms, is that the failure of GNN-based methods under partially observed features is not a reason to discard feature information; rather, the classical label propagation algorithm should be brought back into the architecture. GOODIE propagates labels along graph structure while a separate branch propagates observed features, and a GNN-based decoder forces the label-propagation branch's hidden embeddings to align with the feature-propagation branch's embeddings. A Structure-Feature Attention module then decides how much to trust each source, and pseudo-label contrastive learning refines the pseudo-labels produced by the label-propagation branch. The abstract reports that this hybrid beats current state-of-the-art methods in both low-feature and high-feature settings.
Load-bearing premise
The load-bearing premise is that the GNN decoder can align the label-propagation embeddings with the feature-propagation embeddings without erasing the structural signal that makes label propagation strong when features are scarce; if that alignment is lossy, the hybrid degenerates into one of its branches and the reported gains disappear.
Editorial extensions
If this is right
- If the central claim is correct, a single model can serve both sparse-feature and dense-feature settings, so practitioners would not need to switch methods when feature availability changes.
- It would mean that classical structure-only methods are competitive with, and can beat, GNNs in low-feature regimes when they are properly hybridized with feature propagation.
- The learned attention weights would provide a way to see, for a given graph, how much the model relies on structure versus features, and how that reliance shifts as features are masked.
- The pseudo-label contrastive objective would offer a recipe for improving semi-supervised node classification beyond what either label propagation or feature propagation alone achieves.
Reading between the lines
- The same branch-alignment idea could carry over to other partially observed data tasks, such as link prediction or feature imputation, where a structure-only branch and a feature-only branch could be aligned and then weighted by attention.
- The paper's regime claim implies a useful evaluation practice: report node classification accuracy across a continuous sweep of feature-missing rates rather than at one fixed mask, because the method's advantage is precisely a flatter accuracy curve across the sweep.
- A direct test of the attention mechanism is available: inspect the learned Structure-Feature Attention weights under increasing missing-feature rates, and they should shift toward the structure branch as features become scarcer.
- A stripped-down version of GOODIE without the contrastive objective would reveal whether that component is load-bearing in the scarce-feature regime, a comparison the paper does not explicitly isolate.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The submission, arXiv:2508.01209, is framed as a cs.LG paper titled "Oldie but Goodie: Re-illuminating Label Propagation on Graphs with Partially Observed Features." The abstract claims a novel framework, GOODIE, that combines Label Propagation and Feature Propagation branches via a GNN-based decoder, Structure-Feature Attention, and Pseudo-Label contrastive learning, and that this framework achieves state-of-the-art node classification performance both when few features are available and when features are abundant. However, the full text attached to the submission is a different paper: arXiv:2508.01213, a cs.CL paper by Zhu, Rzeszotarski, and Mimno on modeling the evolution of request-making in human-LLM conversations (ReCCRE). None of the method description, experiments, datasets, baselines, or results for GOODIE are present. The abstract alone provides no quantitative evidence and no way to audit the architecture, the missing-feature protocol, or the comparison protocol. The central SOTA claim is therefore unverifiable from this submission.
Significance. If the claimed results hold, GOODIE would address a genuine and practically relevant gap: GNNs degrade under partially observed node features, and structure-only methods (e.g., label propagation) can outperform them in low-feature regimes. The specific design elements named in the abstract—aligning the Label Propagation branch's hidden embeddings with Feature Propagation embeddings, an attention mechanism over structure and feature information, and pseudo-label contrastive learning—are plausible components that could yield a meaningful contribution. However, because the submission's full text is an unrelated paper, the empirical validation that would justify a SOTA claim is entirely absent. The significance of the work cannot be assessed on the abstract alone, and the paper as submitted does not permit even a preliminary evaluation of the method's novelty or effectiveness.
major comments (3)
- [Full Text (entire submission body)] The document supplied as the full text is not the paper described by the title and abstract. It is arXiv:2508.01213, "Show or Tell? Modeling the Evolution of Request-Making in Human-LLM Conversations" by Zhu, Rzeszotarski, and Mimno. As a result, the GOODIE architecture (GNN-based decoder, Structure-Feature Attention, Pseudo-Label contrastive learning), its training objectives, the experimental setup (datasets, feature missingness rates, baselines, hyperparameters, ablations), and all reported results are missing. The abstract's central claim that GOODIE outperforms existing state-of-the-art methods, in both scarce- and abundant-feature regimes, is entirely unsupported by the supplied text. This is a load-bearing failure: no amount of local revision to the existing text can supply the missing evidence, and the correct manuscript would have to be resubmitted.
- [Abstract (claim of extensive experiments)] The abstract asserts "extensive experiments" and a SOTA result, but even taken by itself it gives no concrete information: no datasets are named, no baseline methods, no feature missingness ranges, no accuracy or error numbers, and no statistical significance measures. For an empirical state-of-the-art claim, these details are the evidentiary core; their absence means the claim cannot be checked that the reported gains are real, consistent, or fairly obtained. This is not a presentation issue but a missing basis for the paper's central assertion.
- [Abstract (Pseudo-Label contrastive learning)] The abstract describes pseudo-labels originating from the LP branch being used for contrastive learning. Without a method section, it is impossible to assess whether this introduces label leakage or a circular training signal, e.g., whether the pseudo-labels of a node are derived from the node's own features or from its neighborhood in a way that the downstream classifier could exploit during training. This is a concrete correctness risk that the manuscript does not address because the relevant derivations and experiments are absent.
minor comments (3)
- [Abstract] The abstract contains several informal phrasings that would benefit from tightening in a revision: "Oldie but Goodie," "Now called by GOODIE," and "a few or the majority of node features" are not standard technical formulations, and the acronym expansion is used inconsistently.
- [Abstract] The statement "Source code of GOODIE is available at: https://github.com/SukwonYun/GOODIE" cannot substitute for the technical description and evaluation that belong in the paper; moreover, a reviewer cannot verify a private repository at review time. The code link may be useful for reproducibility, but it does not repair the missing manuscript content.
- [Limitations (in the attached ReCCRE full text)] The Limitations section in the supplied full text concerns the ReCCRE dataset and human-LLM request analysis; it does not address the GOODIE method or its claims. This is consistent with the document mismatch, and no limitation statement about GOODIE is available in this submission.
Circularity Check
No circular derivation can be identified because the supplied full text is an unrelated paper; the claimed derivation chain is absent, so the SOTA claim is unverifiable but not circular.
full rationale
The abstract of arXiv:2508.01209 makes an empirical SOTA claim for GOODIE, but the supplied 'FULL TEXT' is arXiv:2508.01213, a cs.CL paper on ReCCRE request-making. Consequently there are no equations, no architecture description, no pseudo-label loss definitions, no experimental tables, and no baselines from GOODIE to audit. Circularity analysis requires exhibiting a specific reduction such as Eq. X equal to Eq. Y by construction or a fitted parameter renamed as a prediction; no such step can be quoted because the derivation text is absent. The document mismatch is a completeness failure affecting verifiability, not circularity. Accordingly, no circular step is flagged and the score is 0.
Assumptions & free parameters
free parameters (3)
- Structure-Feature Attention weights
- Pseudo-label contrastive objective hyperparameters
- Propagation iteration counts and decoder depth
assumptions (3)
- domain assumption Feature propagation recovers usable feature estimates under the missingness patterns tested
- domain assumption The benchmark evaluation is fair and representative of real-world partial observability
- domain assumption Pseudo-labels used in contrastive learning do not leak evaluation-label information
Cite this review
Pith. "Pith review of Oldie but Goodie: Re-illuminating Label Propagation on Graphs with Partially Observed Features." pith.science (2026). https://pith.science/paper/3YUESVJP
@misc{pith2026250801209,
author = {Pith},
title = {Pith review of: Oldie but Goodie: Re-illuminating Label Propagation on Graphs with Partially Observed Features},
year = {2026},
howpublished = {\url{https://pith.science/paper/3YUESVJP}},
note = {Machine review of arXiv:2508.01209}
}
read the original abstract
In real-world graphs, we often encounter missing feature situations where a few or the majority of node features, e.g., sensitive information, are missed. In such scenarios, directly utilizing Graph Neural Networks (GNNs) would yield sub-optimal results in downstream tasks such as node classification. Despite the emergence of a few GNN-based methods attempting to mitigate its missing situation, when only a few features are available, they rather perform worse than traditional structure-based models. To this end, we propose a novel framework that further illuminates the potential of classical Label Propagation (Oldie), taking advantage of Feature Propagation, especially when only a partial feature is available. Now called by GOODIE, it takes a hybrid approach to obtain embeddings from the Label Propagation branch and Feature Propagation branch. To do so, we first design a GNN-based decoder that enables the Label Propagation branch to output hidden embeddings that align with those of the FP branch. Then, GOODIE automatically captures the significance of structure and feature information thanks to the newly designed Structure-Feature Attention. Followed by a novel Pseudo-Label contrastive learning that differentiates the contribution of each positive pair within pseudo-labels originating from the LP branch, GOODIE outputs the final prediction for the unlabeled nodes. Through extensive experiments, we demonstrate that our proposed model, GOODIE, outperforms the existing state-of-the-art methods not only when only a few features are available but also in abundantly available situations. Source code of GOODIE is available at: https://github.com/SukwonYun/GOODIE.
Reference graph
Works this paper leans on
-
[2]
Facilitating longitudinal interaction studies of ai systems. InAdjunct Proceedings of the 38th Annual ACM Symposium on User Interface Software and Technology, pages 1–5. Zilin Ma, Yiyang Mei, Krzysztof Z. Gajos, and Ian Arawjo. 2024. Schrödinger’s update: User percep- tions of uncertainties in proprietary large language model updates. InExtended Abstracts...
work page 2024
-
[2022]
Terran Mott, Aaron Fanganello, and Tom Williams
Please be polite: Towards building a politeness adaptive dialogue system for goal-oriented conversa- tions.Neurocomputing, 494:242–254. Terran Mott, Aaron Fanganello, and Tom Williams
-
[2024]
What a thing to say! which linguistic polite- ness strategies should robots use in noncompliance interactions? InProceedings of the 2024 ACM/IEEE International Conference on Human-Robot Interac- tion, pages 501–510. Sheshera Mysore, Debarati Das, Hancheng Cao, and Bahareh Sarrafzadeh. 2025. Prototypical human-ai collaboration behaviors from llm-assisted w...
arXiv 2024
-
[2025]
arXiv preprint arXiv:2504.15236
Values in the wild: Discovering and analyz- ing values in real-world language model interactions. arXiv preprint arXiv:2504.15236. Jonathan Ivey, Shivani Kumar, Jiayu Liu, Hua Shen, Sushrita Rakshit, Rohan Raju, Haotian Zhang, Aparna Ananthasubramaniam, Junghwan Kim, Bowen Yi, and 1 others. 2024. Real or robotic? assessing whether llms accurately simulate...
arXiv 2024
Reviewed August 6, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.