acceptodds
Withdrawn from ICLR 2027

Dual-View Uncertainty Estimation with Noise Elimination for Positive-Unlabeled Graph Classification

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Positive-unlabeled (PU) graph classification aims to learn graph-level classifiers from labeled positive graphs and mixed unlabeled graphs, where only known active compounds are labeled. Due to limited labeled graphs, existing PU graph classification methods mainly focus on pseudo-label generation. However, relying on single prediction confidence may amplify confirmation bias under overconfident mispredictions, leading to noisy graph representations and degraded generalization. To address these issues, we propose Dual-view Uncertainty Estimation with Noise Elimination (called DUNE), which jointly refines pseudo-label reliability and suppresses noisy supervision for robust PU graph learning. Specifically, DUNE first introduces a dual-view uncertainty estimation strategy that evaluates unlabeled graphs from both intra-graph structural consistency and inter-graph distribution discrepancy. It generates perturbed graph views and employs a Pareto-guided selection to obtain reliable pseudo-labels, while constructing a graph-of-graphs structure to capture inter-graph dependencies and calibrate confidence. Furthermore, DUNE employs dynamic noise elimination, where confident pseudo-labeled graphs are re-evaluated according to their representation deviation from class prototypes during later training to remove harmful mislabeled graphs. Extensive experiments on multiple public graph classification benchmarks demonstrate that DUNE consistently outperforms state-of-the-art PU graph classification baselines.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Positive-Unlabeled Learning, Graph Classification, Pseudo-Labeling, Uncertainty Estimation, Noise Elimination.

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_dnw6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper looks at PU graph classification, where only a few graphs are labeled as positive and the rest are unlabeled. PEACE checks whether predictions stay consistent under graph perturbations, uses relations between graphs to help assign labels, and then filters out suspicious pseudo-labels using class prototypes. Experiments on several datasets show how it performs, with ablations to check what each part brings.

Critical strengths
  1. PU graph classification is a meaningful problem, and unreliable pseudo-labels are a real issue in this setting.
  2. The overall pipeline is clear and easy to follow. The teaser also gives a good overview of the main idea.
  3. Experiments cover several datasets, with ablations and sensitivity studies to check how the method behaves.
Critical weaknesses
  1. The method may still introduce errors at several steps. Random edge deletion could remove key structures, and the model might make the same wrong prediction across multiple perturbations. Its performance may also depend on the initial model's ability to separate the classes and whether unlabeled data fits the setting.
  2. These components rely on perturbations preserving key semantics, similar graphs having similar labels, and class prototypes being representative, but these conditions don't always hold. The first two modules rely on the initial model's predictions, while prototype filtering depends on earlier selections, so errors also can carry over and even get amplified. It's still unclear how much the modules correct each other's mistakes or gain from being combined.
  3. The gains seem modest, while the method adds extra computation. There's no trade-off study showing whether the extra cost is worth it, and GDE seems to bring limited gains in the ablation results.
  4. Sec. 3.5 says samples that fail the filtering step aren't used as reliable pseudo-labels, but Eq. (15) still puts the remaining unlabeled samples outside (Ufinal)(U_{\mathrm{final}}) and (W)(W) into the negative-class loss. These two treatments don't seem consistent.
Questions for authors
  1. Do the three modules help correct each other's mistakes, or can errors still carry over from one stage to the next? Does combining them add much beyond using them separately?
  2. How much extra computation does PEACE add compared with the baselines? Is the gain worth the extra cost, especially since GDE seems to bring limited gains on some datasets?
  3. In Eq. (15), why are the remaining unlabeled graphs outside (Ufinal)(U_{\mathrm{final}}) and (W)(W) treated as negatives? Could some still be positive?

If these points are clarified, I'll reconsider my score.

Reviewer_hMKFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PEACE for positive-unlabeled graph classification. It selects pseudo-labels using two views of evidence and then filters them with prototypes. It beats DELTA-PU in most settings. However, key settings are missing, the small gains lack variance estimates, and the filtering step helps only on some datasets.

Critical strengths
  1. The motivation fits the PU setting. When only positive labels are available, wrong pseudo-labels hurt later training. The paper checks each graph's prediction consistency under perturbation, uses evidence from relations between graphs, and then removes suspicious pseudo-labels. Together, these steps form a complete approach to the one-sided supervision problem.
  2. The parts of the method are well thought out. The method has four parts, each using a different source of information (Eqs. 1–14). First, it builds perturbation subgroups at several granularities. Second, it propagates labels on a graph of graphs. Third, a sample is admitted only when both views agree. Finally, it filters samples late in training using leave-one-out prototypes. The leave-one-out step keeps a sample out of the prototype that is used to judge it.
  3. There are clear gains over a strong baseline. Tables 1–2 cover 20 dataset–metric conditions with positive label ratios of 0.1 and 0.2. The method beats DELTA-PU in 17 of them. The average gain is about 1.18 points in Macro-F1 and 0.98 points in Accuracy. Some gains are large; for example, on MUTAG at ratio 0.2, Macro-F1 is 73.82 versus 67.87.
  4. The paper checks the pseudo-labels directly. Besides final metrics, Table 6 reports the error rate, how many wrong and correct pseudo-labels are removed, and the retention ratio. The error rate of kept pseudo-labels goes down on four datasets, and the average drops from 27.4% to 23.3%. This directly supports the filtering step, though only under some conditions.
Critical weaknesses
  1. The novelty is mainly in combining existing ideas, and the theory says little about reliability. Theorem 3.1 describes prediction consistency using the JSD identity and Pinsker's inequality. The authors admit that consistent predictions do not mean correct pseudo-labels, and that errors can remain after the two-view check. So the value of the combined method rests mostly on the experiments.
  2. The initial model and several key settings are not described. Algorithm 1 gives the order of steps: build pseudo-labels, train, filter late, and continue training. Appendix C.3 gives the network, the optimizer, 100 epochs per stage, and some thresholds. What is missing is how the first classifier and representations are obtained, since they are used to build the graph and choose negative anchors. Also missing are the number of views, the set of candidate granularities, the number of repeated groupings, the propagation coefficient, the number of negative anchors, and the filtering percentiles. This makes the results hard to reproduce. To be clear, I am not suggesting that a random classifier was used.
  3. It is unclear whether the small gains are reliable. The main experiments use stratified 10-fold cross-validation, and the authors say all methods use the same splits and label ratios. But the tables do not show the spread across folds or seeds, or intervals for the paired differences. Many gains over DELTA-PU are only a few tenths of a point. From the means alone, we cannot tell whether the advantage is stable. For the same reason, we also cannot conclude that the gains are not significant.
  4. The benefit of GDE depends on the dataset. I compared the 15 conditions in Table 5 one by one. The full model is worse than the version without GDE in 8 of them, and the average Macro-F1 gain is about 0.13 points. In Table 6, GDE removes 41.0% of the correct pseudo-labels on average, and on NCI109 the error rate of kept pseudo-labels does not go down. The authors do discuss this quality–quantity trade-off and the weak effect on some datasets. The open questions are whether prototype filtering is better than a simpler filter at the same retention rate, and how to decide when to use it.
Questions for authors
  1. How are the initial classifier and representations in line 1 of Algorithm 1 obtained? If there is a warm-up, what are its loss, data, and epochs? Please also give the values, and how they were chosen, of M, S, the number of groupings, ζ, the number of negative anchors, and the filtering thresholds.
  2. Can you report paired differences against DELTA-PU on the same folds and seeds with their uncertainty, and say which validation labels and search budget were used for tuning?
  3. At the same retention rate, how does GDE compare with random removal or confidence-based filtering? Can its use or threshold be set on a validation set?
  4. How correlated are the errors of the two views? Also, since Eq. 3 is monotone in both criteria, does the explicit Pareto pre-screening change which candidate is chosen when 0 < λ < 1?
Reviewer_E3HYOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PEACE for positive-unlabeled graph classification. It combines prediction consistency across augmented graph views with label propagation over a graph-of-graphs to construct pseudo-labels. These estimates are integrated through reliability-based selection, followed by prototype-distance filtering to remove potentially incorrect assignments during training. The method is evaluated on five graph-classification benchmarks under different observed-positive ratios, using Macro-F1 and Accuracy, with additional ablation, sensitivity, and pseudo-label-quality analyses.

Critical strengths

The paper directly evaluates pseudo-label quality in addition to downstream classification performance. Table 6 reports retained pseudo-label errors and the removal rates of correct and incorrect assignments. The reduction in retained pseudo-label error on four of five datasets provides useful diagnostic evidence that prototype-based filtering can identify some unreliable supervision.

Critical weaknesses

W1. The graph-specific methodological contribution is insufficiently established, and the comparisons do not adequately isolate it.

The principal components—augmentation-based consistency, nearest-neighbor label propagation, reliability-weighted fusion, and prototype-distance filtering—are broadly applicable techniques. Most operations act on sample-level predictions or embeddings after graph encoding. Although graph augmentations and a GNN provide the graph instantiation, the paper does not sufficiently explain which graph-specific difficulty requires the proposed combination or what substantive methodological advance follows from it.

The theoretical result provides a limited justification for the proposed reliability mechanism. Theorem 3.1 expresses the uncertainty measure through the standard generalized Jensen–Shannon divergence identity and applies Pinsker’s inequality. This characterizes agreement among predictions, but does not establish pseudo-label correctness: consistently incorrect predictions can also have low disagreement. The result therefore offers limited support for the broader evidence-certification motivation.

The paper includes nine baselines, but four primarily vary graph pooling, while the remaining comparisons do not directly establish whether simpler, PU-compatible consistency-based self-training or prototype filtering could achieve similar improvements with the same encoder. Including DELTA-PU is relevant, but more targeted controls are needed to distinguish the proposed contribution from a combination of existing reliability mechanisms.

W2. The additional complexity is not adequately justified by computational analysis or consistently supported by the ablations.

PEACE introduces multiple augmented views, repeated subgroup partitions, Pareto-based granularity selection, nearest-neighbor graph construction, iterative propagation, and multiple training stages. However, the paper does not report computational complexity, runtime, or peak memory relative to the baselines and simplified variants. The implementation description also leaves important settings insufficiently specified, including the number of views, candidate subgroup granularities, and partition repetitions. This makes the practical overhead and scalability difficult to assess.

The ablations do not demonstrate consistently beneficial contributions from every component. In Table 5, removing GDE improves performance in eight of the fifteen dataset–ratio settings; averaging those entries gives the full model an advantage of only approximately 0.13 percentage points. For example, removing GDE improves Macro-F1 from 83.50 to 85.55 on BAMultiShapes at an observed-positive ratio of 0.1. These results weaken the justification for including the complete pipeline rather than a simpler variant.

The main and ablation tables do not report standard deviations, confidence intervals, or paired statistical comparisons, despite using ten-fold splits. Consequently, the robustness of small improvements remains unclear—for example, the Macro-F1 difference over DELTA-PU on NCI109 at an observed-positive ratio of 0.2 is only 0.02 percentage points. Table 6 provides useful evidence about pseudo-label quality, but does not resolve whether filtering yields a reliable downstream classification benefit sufficient to justify its added complexity.

Questions for authors

Which technical aspects of PEACE specifically address difficulties unique to graph-level PU learning, beyond applying general consistency estimation, sample-level label propagation, and prototype filtering to graph embeddings? How would the method compare with simpler PU-compatible combinations of these mechanisms?

What are the training-time and memory overheads relative to DELTA-PU and the simplified PEACE variants? Please clarify the number of augmented views, subgroup granularities, partition repetitions, and propagation stopping criterion needed to reproduce and assess these costs.

Are the reported improvements robust across paired folds or independent runs? In particular, how should the frequent improvements from removing GDE be reconciled with the claim that all components contribute to the overall performance?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.