acceptodds
Withdrawn from ICLR 2027

SymRankor: Symmetric Pairwise Ordinal Learning for Semi-Supervised Regression

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Continuous-valued labels are often costly to obtain, motivating semi-supervised regression methods that learn from abundant unlabeled data. Confidence-based classification methods do not transfer directly, because a scalar regression output provides no normalized class confidence for selecting pseudo-labels. State-of-the-art methods such as RankUp address this mismatch with an auxiliary ranking classifier (ARC) over sample pairs, but its unlabeled pairwise objective uses cross-entropy and does not explicitly account for corrupted pseudo-orders. We propose SymRankor, which treats the weak-view order prediction as a potentially corrupted pairwise pseudo-label under a Bradley–Terry interpretation and applies a symmetric loss to the strong-view score difference. Under a theoretical label-flip model, the unlabeled-pair risk is a positive affine transform of its clean counterpart. This transformation preserves risk ordering and any attained argmin set without estimating the flip rate. Experiments across vision, audio, and text show that SymRankor improves upon its direct predecessor, with the clearest gains on audio and text. The symmetric objective remains competitive across the evaluated modalities even without confidence filtering. This study explicitly examines semi-supervised regression from this perspective. A chemical case study further suggests that the same ordinal objective can exploit coarse domain-derived ordinal information.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Weak supervision, Learning to rank, semi-supervised regression

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_gA2COpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper builds on RankUp and proposes SymRankor, which treats self-generated pairwise orders as noisy pseudo-labels and replaces the cross-entropy loss on the unlabeled ranking term with a symmetric loss to improve robustness against noisy pseudo-labels. The method is evaluated on vision, audio, and text benchmarks, with an additional chemistry case study that incorporates coarse ordinal information derived from domain knowledge.

Critical strengths
  1. Replacing the cross-entropy loss on the unlabeled ranking term with a symmetric loss is a lightweight change to RankUp that could be easily adopted.
Critical weaknesses
  1. The effectiveness of the symmetric loss is not convincingly demonstrated. The primary SymRankor results in Table 1 rely on hyperparameter optimization (HPO), while all baselines, including RankUp, use their default configurations. The tuned hyperparameters include those shared with RankUp (e.g., the ARC loss weights and unlabeled ratio). Without HPO, SymRankor does not outperform RankUp on vision (7.060 vs. 7.026) or audio (0.454 vs. 0.458), and even with HPO it remains behind RankUp + RDA and SSR-RCUS on vision. Since the tuned values vary widely across tasks and no sensitivity analysis is provided, the reported gains may depend more on tuning than on the symmetric loss itself.
  2. Limited novelty and weak theoretical support. The method largely transfers the symmetric pairwise loss of SymPO to RankUp and is evaluated only within RankUp. Its theoretical analysis restates a known result (Charoenphakdee et al., 2019) under a constant flip-rate assumption, which may not reflect the instance-dependent errors of self-generated pseudo-orders. Recent work that explicitly models such instance-dependent pseudo-label noise in semi-supervised regression (Sun et al., NeurIPS 2025) is neither discussed nor compared. Sun, Xueqing, et al. "Semi-Supervised Regression with Heteroscedastic Pseudo-Labels." Advances in Neural Information Processing Systems (2025).
Questions for authors
  1. Could the authors tune RankUp with the same HPO protocol? If SymRankor still outperforms it, this would substantially strengthen the main claim.
  2. How sensitive is SymRankor to its hyperparameters (e.g., β and the ARC loss weights)? A sensitivity analysis would help assess its practicality, given the large variation of tuned values across tasks.
  3. Can the symmetric loss be applied to other semi-supervised regression methods beyond RankUp? Demonstrating this would support its generality.
  4. Could the authors discuss or compare with Sun et al. (NeurIPS 2025), which models instance-dependent pseudo-label noise in semi-supervised regression?
  5. Why is the sigmoid loss chosen as the default? Under the default vision setting with τ\tau = 0, several symmetric losses in Table 13 (e.g., Gaussian, smooth ramp) achieve lower MAE than sigmoid in Table 1.
Reviewer_Sg43OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SymRankor modifies RankUp by replacing the cross-entropy loss on unlabeled pairwise pseudo-orders with a symmetric margin loss. The paper motivates this change through a noise-tolerance identity and evaluates it on vision, audio, and text regression. A molecular-solubility case study additionally incorporates descriptor-derived ordinal supervision.

Critical strengths

The modification is simple and easy to incorporate into an existing semi-supervised regression system. Reducing dependence on confidence filtering is practically useful.

The paper reports both default and tuned configurations and includes a loss-by-threshold ablation. These comparisons help distinguish the effect of the objective from the effect of filtering.

Critical weaknesses

The strongest audio and vision gains depend substantially on hyperparameter optimization. Without tuning, audio performance is close to RankUp, while vision performance is slightly worse. The cross-entropy ablation uses hyperparameters selected for the symmetric objective, rather than its own optimized configuration. Giving both objectives comparable search budgets would make the comparison more convincing. The search budget itself should also be reported.

The risk identity is valid under the stated constant-flip model, but its connection to actual pseudo-label errors remains largely untested. Those errors are generated by the evolving model and may depend strongly on target gaps, confidence, and training stage. Appendix C appropriately limits the identity to the auxiliary term; empirical measurements of the actual noise would better support the proposed explanation.

Appendix D incorrectly states that tied labeled pairs with uniform targets contribute a constant cross-entropy loss. For distinct samples with equal labels, this loss generally pushes the predicted score difference toward zero. Only self-pairs have identically zero gradient. Since these datasets contain many ties, the implementation and its description should be clarified.

Questions for authors

What happens when RankUp and the cross-entropy variants receive the same hyperparameter-search budget as SymRankor?

How do pseudo-order error rates vary with target separation and training progress?

How are tied pairs handled in the implementation, and does correcting the stated interpretation affect any conclusions?

Reviewer_B5j5OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes SymRankor, a semi-supervised regression method. The proposed algorithm is based on RankUp but introduces a symmetric margin loss which limits the impact of undesirable gradients from mis-predicted pseudo labels. The experimental results on ordinal regression benchmarks including vision, text, audio data, and chemical case, show that the proposed algorithm can achieve comparable or better performance than existing methods.

Critical strengths
  • The paper is clearly written and provides a detailed description of the method and theoretical assumptions.
  • The proposed symmetric loss is a simple and practically useful modification for semi-supervised ordinal regression studies.
  • Experiments across vision, audio, and text demonstrate that the method is applicable to different modalities.
Critical weaknesses
  • The experimental performance gain over existing methods is relatively incremental, as the primary algorithmic change from RankUp is replacing the unlabeled cross-entropy loss. More comparison results would strengthen the paper.
Questions for authors
  1. Although the main idea behind the proposed method appears potentially useful, its empirical improvement seems relatively incremental. In particular, on the UTK dataset, the proposed algorithm does not outperform the existing methods. Many other age-estimation datasets are available, including MORPH II, CACD, CLAP, and Adience, as well as related ordinal regression tasks such as dating historical color images (ECCV 2012) and image-aesthetics prediction (e.g., AADB). A broader evaluation on these datasets and tasks would provide stronger evidence for the effectiveness and generalizability of the proposed method.

  2. The proposed symmetric loss appears to introduce a trade-off: it limits the damage caused by incorrect pseudo-orders, but may also weaken the learning signal from reliable pseudo-orders compared with cross-entropy. More importantly, SymRankor does not explicitly model the probability that an individual pseudo-order is correct. A key characteristic of ordinal data is that pairs with nearby ranks are generally more difficult to order correctly, whereas pairs separated by a larger rank gap tend to be more reliable. Would it be more appropriate to incorporate such pair-specific reliability directly into the loss, such that reliable pairs receive a stronger learning signal while uncertain pairs receive the more robust symmetric objective? The current confidence threshold only performs hard selection and treats all retained pairs equally.

  3. A t-SNE visualization of the learned representations over training would be interesting and informative. It would also be useful to examine whether similar representation patterns emerge across different modalities. If possible, visualizing both labeled and unlabeled samples would provide further insight into how the proposed algorithm organizes the representation space.

  4. [Minor] It would be good to have a brief discussion of related order-learning approaches such as Chainization (ECCV 2022), since it also utilizes pseudo-pairs derived from partially ordered data in a semi-supervised setting.

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

Related papers

Discussion (0)

Sign in to comment.