acceptodds
Under review as a conference paper at ICLR 2027

Correcting Selection Bias in Positive and Unlabeled Generation with Selection Records

Abstract

Positive and unlabeled (PU) data combine labeled positives with a population sample whose classes are unknown. A generator fitted to the labeled positives inherits their selection bias: positives that are verified more often stay overrepresented, even when every label is correct. We correct this bias with selection records, which show whether objects in a population sample were selected for verification but not their classes. Under shared selection and constant reporting, the normalized density ratio of labeled positives to selected objects identifies the positive distribution without selection probabilities or the population frequency of positive labels. We bound the resulting distribution error by the ratio estimation error and guide a diffusion model with the fitted weights. Under simulated biased selection on MNIST, CelebA and CIFAR-10, the weights improve the group composition of the weighted positives, and joint class and group error falls on MNIST and CIFAR-10. On MNIST, generated composition also improves at matched fractions of predicted positives, although even weights equal to the true class leave generation error. On CelebA, longer ratio training lowers the joint error of the weighted target, but the held out classification loss does not select it. Further experiments separate the information in the records from their number and show that restricted probability models, mismatched records and class dependent selection prevent recovery.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
positive-unlabeled learning, classifier guidance, diffusion models, density ratio estimation, exponential tilting, weak supervision

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_7uEWOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper deals with biased selection PU model under case control scenario, when additional information is given whether the item will undergo possible labeling or not. Namely, labeling proces itself is proceded by verification, only observations which get positive verification status (A=1) are potentially labeled under SCAR i.e. positive observations are labeled with constant probability c, and the remaining verified positives and all verified negatives are unlabeled. The observer is given a sample L of labeled observations, sample U of observations without class indicators pertaining to general population and a sample V of observations without class indicators but with corresponding information whether particular item was verified or not. Thus, the fact that labeled observations do not follow positive class distribution, as in SCAR case, is due to the fact that verification depends on predictors. The aim is to generate new observations from positive class by adapting diffusion process approach using available data. This is based on estimation of density ratio rAr_A of densities of positively verified items to density of labeled items and the main theoretical result (Theorem 2) establishes bound on total variation distance between positive class distribution and r^A\hat r_A-weighted distribution of general population. Experiments show usefulness of the approach in generating items from the positive class over methods which disregard inherent bias of labeled observations.

Critical strengths

The approach nicely disentangles process which leads to biasedness of labeled observations into verification process and labeling process itself. By using conditional independence of verification and class indicatator given predictors (Y⊥A∣XY\perp A |X) (which is not true when A is replaced by labeling indicator S), this allows for estimation of density ratio rAr_A and, consequently, distribution of positive class.

Critical weaknesses

The main weakness is that under conditional independence assumption and with observability of V sample, the problem becomes conceptually rather simple. Namely, in the present setting propensity score e(x)e(x) is proportional to posterior probability v(x)v(x) of being positively verified, and as V sample is available, v(x)v(x) is estimable, while estimation of the proportionality constant can be avoided by suitable normalization. Moreover, crucial ratio r=f+/fr =f_+/f can be estimated via estimation of rAr_A. The second problem is that in the present format the paper is very hard to read, with essential information mostly contained in the appendix. In particular, the algorithms for the method are not given in the main body of the paper, but only on p. 21 in the appendix, and the essence of Inverse Probability Weighting ( IW) and Selected Reference (SW) approach, which is crucial, is also given there (appendix B2). In my opinion the paper requires substantial restructuring.

Questions for authors
  1. Estimation of rAr_A can be performed using many alternative methods such as direct density-ratio estimation methods e.g. relative unconstrained least-squares importance fitting (Yamada et al. 2012). Why the method employed in the paper is preferable ?
3 comments

Though the presentation in our initial submission was not really easy to follow, we are very grateful for your careful review and deep understanding of our method. As you point out, the cancellation is straightforward under our assumptions, but the main text made the method harder to follow than necessary. We have therefore reorganized the paper around the data flow and sharpened the statement of our contribution.

1. The method now reads from the main text. Sections 2–3 follow the data through identification, weight estimation and generation.

  • Section 2 defines the three observed samples and separates selection from positive reporting. Figure 2 shows why changing the reference cancels the shared selection factor.
  • Section 3.1 presents SR and IPW side by side: SR compares labeled positives with all selected objects; IPW estimates selection probabilities and reweights labeled positives. It also lists the SR steps in order: source classification, exponentiation, mean normalization on U, and truncation. Source labels mark sample origin, not class.
  • Section 3.3 and Figure 3 connect the weights to guidance training and sampling. Clean-image weights stay fixed as noise is added; at sampling, the potential's image gradient corrects the base score, and both networks stay frozen.

Detailed objectives and pseudocode are in Appendix C.1. Sections 4.1–4.4 then examine record information, weighting strategies, generation and assumption violations in turn.

2. What the simple identity buys. A generator can fit labeled positives perfectly and still reproduce the wrong proportions within the positive class. The identity says which reference fixes this. Under shared selection, constant positive reporting and overlap, rA=fl/fAr_A=f_l/f_A, where fAf_A includes both classes, and (rA/EPrA)P=P+(r_A/\mathbb E_P r_A)P=P_+. SR estimates these corrective weights without a propensity model.

The identity leaves three practical questions open, and the paper answers each. How does ratio error affect the target? Theorem 2 bounds the target error by the relative ratio error for any nonnegative estimator, and Section 3.3 separates it from the error of generating from the target. Do the records carry information beyond reference size? With equal reference counts, selected rather than random references raise weighted-resampling coverage by 2.89 ± 0.70 percentage points on MNIST under strong selection and by 1.86 ± 0.30 on biased CelebA, at a CelebA precision cost of 1.60 ± 0.49 (five paired repetitions; SE); constant selection shows no comparable gain (Section 4.1). Does the correction reach generated samples? On MNIST, SR lowers intensity-group TV at matched predicted purity; on CelebA it improves FID but worsens joint composition (Section 4.3).

As you note, normalized IPW also removes the unknown scale, and in our matched MNIST controls a correctly specified IPW model performs within one paired SE of SR (Section 4.2). SR gets there without specifying a selection model. Kato et al. (2023) used exposure records for PU classification, and conditional energy guidance is established. Our contribution is the distribution-recovery formulation for generation, its estimator and error analysis, and the controlled study of when the correction helps.

3. Why logistic estimation? Two reasons. It is a direct ratio estimator that fits naturally on image networks: with balanced source classes, the population-optimal logit is log⁡(fl/fA)\log(f_l/f_A), so exponentiation gives the ratio without estimating either density. Its excess population logistic loss also controls the target-distribution error when ratios are bounded and selection probabilities are bounded away from zero (Appendix B.2.1).

RuLSIF is a natural alternative and slots into the same pipeline: Theorem 2 holds for any nonnegative estimator, and Appendix B.2.3 shows how relative ratios map to the ordinary ratio, including when direct normalization preserves the target. In matched MNIST controls with the same network and a fixed 200-update budget, neural relative least-squares estimates have higher joint class-by-intensity TV than logistic SR in every tested setting, though one gives better within-positive composition in one setting (Appendix E.2.2). This is a neural adaptation rather than kernel RuLSIF with its own model selection, so it shows that logistic is a sound choice, not that it is the best one.

The Authors have improved the readability of the manuscript in its revised version, which is now much easier to follow. But I believe alternatives to estimation of density ratio by minimising the balanced risk pertaining to log-loss should be explored more thoroughly. Indeed, the method applied is essentially classification based approach to estimate the ratio, and this is known to work poorly for high dimensions and/or well separated densities. Still, I am inclined to give the Authors the benefit of doubt. I retain my score.

Thank you for recognizing the improved readability and for pressing us on density-ratio estimation. We agree that our earlier 200-update neural relative-loss control was not a sufficient comparison. New Appendix G, referenced from Sections 4.2 and 4.3, compares kernel uLSIF and RuLSIF with logistic estimators on CelebA64, first for weighted targets and then for generated images.

Setup. All estimators use the same 512-dimensional frozen CLIP features, five source blocks and 100,000 images in U: linear logistic, Gaussian-kernel logistic (sharing the least-squares basis, to separate loss from basis), uLSIF, and RuLSIF with ρ = 0.1 and 0.5. Each selects hyperparameters by its own validation loss on three folds of the source fitting data; neither target attributes nor target TV enter any choice. After source validation favored grid boundaries, we expanded the search to 100–500 centers, bandwidth factors 0.25–4 and ridge values 1e-6 to 1.

Weighted targets. Joint Smiling×Male TV×100 of SR's continuous weights to the fixed positive reference (predicted attributes; lower is better), with ESS as % of U; means ± descriptive SE over five reused blocks (Table 38):

EstimatorTarget TVESS (%)
Previous linear logistic19.65 ± 0.3869.71 ± 0.84
Updated linear logistic8.53 ± 0.3846.33 ± 0.77
Kernel logistic15.65 ± 0.8457.29 ± 1.60
uLSIF19.89 ± 0.4968.78 ± 0.68
RuLSIF ρ=0.1, direct19.42 ± 0.7668.71 ± 1.21
RuLSIF ρ=0.1, inverse17.72 ± 0.7565.49 ± 1.20
RuLSIF ρ=0.5, direct23.22 ± 0.4776.56 ± 0.90
RuLSIF ρ=0.5, inverse11.72 ± 0.6748.89 ± 2.54

First, the comparison exposed a consequential choice in our own estimator: all ten updated linear fits select ridge 0.1, missing from our previous grid, which lowers SR target TV by 11.12 ± 0.10 (paired SE) in every block. Second, the ranking depends on the reference. Held-out source AUC is 0.79–0.83 against all records but 0.68–0.73 against selected records. With the more separable all-record reference, every least-squares rule (18.04–23.63) beats updated linear and kernel logistic (24.64 and 24.52; Table 39), consistent with your point about well-separated sources, although the two references define different targets. Better source log loss alone does not give a better correction. Third, with SR's fixed source threshold, RuLSIF ρ=0.5 (4.18 ± 0.71) and updated logistic (4.32 ± 0.51) come close to, but do not improve on, the existing 3.80 ± 0.55.

Generated images. With the diffusion base frozen, we trained 70 new guidance potentials (width 96, 20,000 updates; 200-step sampling; 1,024 images per block) and reused the 25 existing controls. Kernel runs use the initial source grid, so their targets differ from the table above: uLSIF and RuLSIF ρ=0.1 with both references and all weight transformations. The linear follow-up was motivated by the target results. Paired changes in generated outcomes (mean ± paired SE over the same blocks; coverage at k=5 in pp; Tables 40–42):

ComparisonJoint TV×100FIDCoverage
Updated − previous SR soft−4.79 ± 0.45−0.24 ± 0.11+0.92 ± 0.20
Updated SR soft − previous SR threshold+3.36 ± 1.04+0.11 ± 0.14+0.25 ± 0.19
RuLSIF inverse SR − previous SR threshold+8.59 ± 1.22+0.34 ± 0.08−1.15 ± 0.38

The estimator improvement reaches generation: joint TV and coverage improve in all five blocks. The existing SR threshold still gives the lowest joint TV (18.52 ± 0.85); every kernel method is worse than it in every block, and refitting the thresholded linear estimate gives 18.96 ± 1.23. With both linear soft fits updated, SR improves FID by 2.93 ± 0.32 and coverage by 1.95 ± 0.38 pp over the all-record reference but raises joint TV by 2.50 ± 0.62. Target gains also outpace output gains: the updated soft target predicts 78.19% Smiling but generation reaches 64.84%, and its TV to its own target rises from 7.07 to 13.36 (Table 43).

Result. 204 of 1,050 kernel-logistic candidates had an unconverged fold, mainly at small ridge, and were excluded; all selected fits converge, and least-squares fits are direct linear solves. Some selected kernels sit at grid boundaries, so these results hold within the stated ranges. Within them no estimator wins across references and weight rules: on high-dimensional features, loss, regularization, reference and weight transformation need to be examined together.

Thank you again; this suggestion improved our own estimator in both targets and generated images.

Reviewer_k1w6OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper tries to generate the full positive class when the labeled examples are skewed by who gets checked. It uses selection records, which only say whether an image was picked for checking, not its class, to reweight a box of unlabeled images and guide a diffusion model. They test this on MNIST (even digits), CelebA (smiling faces), and CIFAR-10 (vehicles), with fake selection based on brightness or the Male label. The reweighting does fix the mix of groups. The generated pictures only clearly improve on MNIST. On CelebA and CIFAR-10 the score barely moves for generation. The generation improvements are nominal/statistical noise, while the headline suggests it improves generation.

Critical strengths
  • This is a well-motivated and thoughtfully executed study that introduces a simple and compelling idea: selection records can correct PU selection bias for generation by canceling the selection mechanism in the ratio of labeled positives to selected objects.
  • Integrates the corrected target with diffusion generation through conditional energy guidance, preserving the correction by fitting weights on clean images and propagating them across noise levels.
  • Honest separation of weight error vs generation error (true-class weights still leave a generation gap)
  • Group TV improves under biased selection on all three datasets.
Critical weaknesses
  • Generation story oversold relative to results, where the clear win was mainly on MNIST but CIFAR and CelebA coverage gains are nominal
  • Comparisons to recent debiasing methods for diffusion training (e.g., training unbiased diffusion models or likelihood-free importance weighting approaches) are not run empirically, even though related work is cited.
  • Empirical comparisons to debiased diffusion training are somewhat limited.
Questions for authors
  • Why is main Table 1 only k=5 when you already have k=3,5,10, and will you drop bold marks when gaps are within 1 SE?
  • With 10+ repeats that redraw L, V, and U, do MNIST coverage gains and RGB “wins” still hold?
  • How does SR compare to a thresholded PU classifier with uniform weights over predicted positives on the image experiments?
3 comments

Thank you for the careful assessment. We agree that the original RGB generation gains were nominal. Below we answer each weakness and question in turn.

1. Generation claims. With the original soft weights, RGB coverage stays within 0.5 pp of Uncorrected (Section 4.3), and we have no new CIFAR-10 generation result. The abstract and introduction now claim generation gains only for MNIST composition and CelebA64. On MNIST under strong selection, SR lowers intensity-group TV×100 by 14.09 ± 2.95 at matched predicted purity. On CelebA64, SR keeps images whose balanced source logit is at least 0 (fixed before target evaluation), with CLIP features and diffusion base frozen. Relative to Uncorrected soft, joint Smiling×Male TV×100 falls by 6.39 ± 1.55 and FID by 2.97 ± 0.23, while coverage at k=5 rises by 2.21 ± 0.45 pp, each in all five development blocks (Table 3). Most of the joint-TV gain comes from recovering positive mass. These gains replicate on ten fresh blocks (point 4). Stronger guidance, which leaves the threshold target unchanged, further lowers joint TV×100 from 18.52 to 4.30 with unchanged FID (Appendix G.4); same-weight resampling still has lower FID and higher coverage than generation.

2. TIW and LFIW. We have not run these comparisons. TIW (Kim et al., 2024, Section 2.1) needs biased data plus representative target samples, and LFIW (Grover et al., 2019, Section 3) learns a ratio between target and generated samples. Here L is biased, U is mixed and V carries no class outcomes, so the positive target both methods need samples from is precisely what SR estimates. As Reviewer jttG wrote, "to the best of my knowledge, this is the first work to extend selection bias correction to data generation in this setting." Giving these methods true positives would change the problem. Using our weighted U as their reference would be a complementary study that remains open (Related work). We claim no advantage over TIW or LFIW.

3. Neighborhoods and rankings. The main table uses k=5, following Naeem et al. (2020); Tables 22 and 31 report k=3, 5 and 10. The MNIST coverage/precision tradeoff under strong selection holds at all three, and the new CelebA coverage gain is positive at each. The small coverage differences of the original soft weights on RGB change sign across k (Table 22). We removed all bold and underlined mean rankings from Table 2; 6 of its 12 top-versus-second gaps were within one paired SE.

4. Redrawing L, V and U. On CelebA64, the generation gains hold. We repeated the comparison on ten new blocks that jointly redraw L, V and U from the training split, refitting every estimator, with base, guidance and references fixed and methods declared before generation (Appendix G.5, Table 45). Uncorrected threshold applies the same zero-logit cut to L against all record images. Entries are SR threshold minus each comparator (mean ± paired SE; blocks favoring SR):

ComparatorJoint TV×100FIDCoverage, ppMale TV×100 among smiling
Uncorrected soft−4.99 ± 0.88 (9/10)−2.98 ± 0.19 (10/10)+1.02 ± 0.47 (8/10)−2.56 ± 1.10 (9/10)
Uncorrected threshold−2.81 ± 0.79 (8/10)−3.80 ± 0.19 (10/10)+1.76 ± 0.41 (9/10)−3.87 ± 0.91 (9/10)
VPU−2.85 ± 1.06 (8/10)−4.68 ± 0.40 (10/10)−0.34 ± 0.31 (4/10)−5.16 ± 0.96 (10/10)

The coverage gain is smaller than on the development blocks, recall falls by 1.16 ± 0.38 pp, and each U shares about 61% of its images with the original U because all come from one training split. On MNIST, ten joint redraws exist only at the weighted-target level (Table 11): under strong selection SR beats the uncorrected Raw score at every retention fraction in all ten blocks (5.29 versus 9.49 joint TV×100 at 50%), with no consistent difference under constant selection. Five further MNIST L/V draws with U fixed give coverage +3.18 ± 0.90 pp and precision −2.07 ± 1.04 pp at k=5 (Table 24).

5. Thresholded PU classifiers. VPU uses only L and U and weights uniformly the images scoring at least half its maximum. On MNIST weighted targets it beats soft SR under constant selection (joint TV×100 4.72 versus 10.56) but distorts composition under strong selection (32.30 versus 10.63); RGB targets show the same split (Table 1; Appendix E.2.2). In generation, the table above compares VPU with SR threshold on the ten fresh CelebA64 blocks; the five development blocks give the same ordering. VPU images are more precise (+0.72 ± 0.18 pp) but have lower recall (−2.40 ± 0.65 pp), and the distortion is already in VPU's target (joint TV×100 25.81 versus 4.03 for SR). VPU uses its original pixel classifier, whereas SR uses CLIP features. Soft weights are not uniformly best; all four predeclared SR fractions are reported.

I am sorry for my late response. Thank you for your efforts and detailed rebuttal. Overall, I believe the proposed approach of generating the full positive class when labeled examples are skewed could make a valuable contribution to the research community. I would like to change my score from "2: Weak rejection" to "3: Weak acceptance."

Thank you very much for your response and for recognizing the potential contribution of our work.

We noticed that the overall score in the review form may not yet have been updated. When convenient, could you kindly check and update it accordingly?

Thank you again for your time, constructive feedback, and engagement with our rebuttal.

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

This paper proposes a generative method that corrects selection bias in the biased positive-unlabeled (PU) learning setting. In addition to labeled positive data LL and unlabeled data UU, the proposed method uses selection records VV, which indicate whether each data point was selected for label verification. Under certain assumptions, the density ratio between the true positive distribution and the overall data distribution is proportional to the density ratio between the labeled positive distribution and the distribution of all selected data points. Based on this relationship, the proposed method estimates the density ratio using a discriminator. The estimated density ratio is then used to train a separate guidance network, which guides the sampling process of an unconditional diffusion model to correct selection bias in the generated data. Experimental results demonstrate that the proposed method successfully corrects selection bias on MNIST, while improvements in generated images are limited on CIFAR-10 and no clear improvement is observed on CelebA.

Critical strengths

While there are many existing studies on biased PU learning that use propensity scores, to the best of my knowledge, this is the first work to extend selection bias correction to data generation in this setting. I also appreciate that this paper incorporates the estimated density ratio into a guidance network to perform actual data generation and evaluate both the bias and quality of the generated samples.

Critical weaknesses

It would be encouraging to see stronger improvements in the generated images on more realistic datasets such as CelebA. I believe the value of the proposed method would be significantly enhanced if it could be made effective on such complex datasets. Do the authors have any ideas for improving its performance in these settings?

For example, given that the method works well on relatively low-dimensional MNIST but struggles on higher-dimensional CelebA, one possible cause may be the performance of the density ratio estimator. Density ratio estimation can be challenging and numerically unstable, particularly when the estimated ratios take extremely large values. One possible approach might be to use relative density ratios, which are bounded by construction.

Alternatively, the density ratio estimator may be working reasonably well, while the difficulty lies in the diffusion sampling process. For example, the diffusion process on CelebA may exhibit more complex phase-transition behavior than on MNIST, making it more difficult for the guidance network to accurately reproduce the desired distribution.

Questions for authors

Please see Critical Weaknesses.

3 comments

Thank you for separating ratio estimation from diffusion sampling. Following both leads gave us two improvements on CelebA64, one on each side. With the pretrained diffusion base frozen, SR with a fixed source threshold improves generated images over both uncorrected weightings and over soft SR. Entries are SR threshold minus each comparator at the original guidance strength (mean ± paired SE over five source blocks, 1,024 images each; Table 3 and Table 30):

ComparatorJoint Smiling×Male TV×100FIDCoverage at k=5, pp
Uncorrected soft−6.39 ± 1.55−2.97 ± 0.23+2.21 ± 0.45
Uncorrected threshold−4.20 ± 1.27−3.78 ± 0.37+2.36 ± 0.45
SR soft−8.14 ± 1.05−0.35 ± 0.12+0.66 ± 0.14

Every difference favors SR threshold in all five blocks. Stronger guidance then removes most of the remaining attribute gap, lowering joint TV×100 from 18.52 to 4.30 (sampling paragraph below). The same five blocks were used to develop the recipe, so these SEs are descriptive; a replication on ten fresh joint draws of L, V and U is running.

On the ratio side, we first checked your concern about extreme values. The original CelebA weights were not extreme: their mean ESS is 74.85% of U and none reaches the cap of 16 (Appendix E.2.1). The soft target was simply inaccurate, and our earlier soft-SR runs lowered FID but worsened joint TV (Appendix E.3.3). We then tried bounded relative ratios directly with kernel RuLSIF, together with uLSIF and kernel logistic, all on frozen CLIP features and tuned by source validation alone (Appendix G). Bounding did not help. What helped was a better regularized logistic ratio and, above all, the threshold: keep the images that the source classifier finds more like labeled positives than like selected records, and weight them equally. Joint TV×100 of the weighted target and of the generated images (Tables 38–43):

SR weightsTargetGenerated
Previous soft logistic19.65 ± 0.3826.66 ± 0.53
RuLSIF ρ=0.1, bounded ratio21.43 ± 0.9128.24 ± 0.48
RuLSIF ρ=0.1, mapped back19.75 ± 0.9427.11 ± 0.55
Soft logistic, ridge 0.18.53 ± 0.3821.88 ± 0.67
Fixed source threshold3.80 ± 0.5518.52 ± 0.85

A wider kernel search gives bounded RuLSIF targets of 19.42 (ρ=0.1) and 23.22 (ρ=0.5), still no better. Refitting soft logistic lowers generated TV by 4.79 ± 0.45 in every block, yet it and every kernel method remain worse than the threshold in every block.

On the sampling side, your hypothesis was right. At the original strength, the threshold target predicts 89.12% Smiling but generation reaches 68.20%, and resampling U with the same weights reaches joint TV 4.55 instead of 18.52. We resampled the trained potentials of all five blocks, without retraining, with stronger guidance or with guidance limited to one third of the reverse process. Threshold weights take only two values, so stronger guidance leaves the clean target unchanged. All settings are reported (Appendix G.4, Table 44):

SR threshold guidanceJoint TV×100TV×100 to own targetFIDCoverage, %
Strength 118.52 ± 0.8520.91 ± 0.5324.26 ± 0.1689.38 ± 0.48
Strength 1.59.84 ± 1.1612.18 ± 1.0324.20 ± 0.2090.14 ± 0.30
Strength 24.30 ± 1.105.79 ± 1.1524.35 ± 0.2390.72 ± 0.40
Strength 34.98 ± 1.053.68 ± 0.7125.49 ± 0.3391.11 ± 0.42
Strength 2, middle third only15.20 ± 1.1317.53 ± 0.7723.82 ± 0.1389.63 ± 0.39

At strength 2, joint TV falls by 14.22 ± 0.41 in every block, FID is unchanged and coverage rises by 1.35 ± 0.28 pp, so the attribute mix now matches resampling. Two costs remain: recall drops by 1.97 pp, and FID is still 7.04 above resampling. Strength 3 overshoots Smiling and costs 1.24 in FID and 7.03 pp in recall. The same strengths applied to Uncorrected threshold amplify its shortage of Male faces and raise FID by 4.71 at strength 2, so SR's lead widens to 10.86 ± 1.88 in joint TV and 8.40 ± 0.81 in FID. On timing, strength-2 guidance in the middle third alone already beats strength 1 throughout, while the low- or high-noise third alone loses most of the effect. Smiling is thus decided mainly at intermediate noise, in line with your phase-transition picture, but guidance over the whole process is still needed.

Thank you for the additional experiments and detailed analysis. Although the results show improvements, I remain concerned about the quality of the generated images on CelebA. I appreciate your efforts and believe the approach has potential for further improvement. However, these remaining concerns lead me to maintain my rating of Weak Rejection.

Thank you very much for your careful evaluation and constructive feedback. We understand your remaining concerns regarding the quality of the generated images on CelebA and respect your decision to maintain the current rating.

Your comments provide clear and valuable guidance for improving our work during this period. In particular, they will help us focus our subsequent experiments on further improving the generation quality on CelebA and refining our experimental design accordingly. If possible, we will also consider including representative generated images in the final version so that readers can more directly assess the visual quality of our results.

We again sincerely appreciate your time and thoughtful engagement.

Comments on the paper

1 comment

Dear Area Chair,

At the close of the rebuttal period, we thank the reviewers for their careful reading and offer a brief summary of the revisions we made in response to their feedback.

Reviewer feedback. Reviewer jttG wrote that, to their knowledge, this is the first work to extend selection bias correction to generation in this setting. Reviewer 7uEW found the revision "much easier to follow" and maintained a rating of Weak Acceptance. Reviewer k1w6 wrote that the approach "could make a valuable contribution to the research community" and raised their rating from Weak Rejection to Weak Acceptance.

Responses to the main concerns

  1. Readability (7uEW). Sections 2 and 3 now present the method from identification of the positive target through weight estimation and guided sampling. Appendix C.1 gives the full algorithms.
  2. Generation on CelebA64 (k1w6, jttG). Thresholded selected reference (SR) gives equal weight to unlabeled images whose source logit is at least zero. After developing the approach on five blocks (Table 3), we fixed the method and primary comparison and tested it on ten new joint draws of labeled positives, selection records, and unlabeled data from the same training split. With the diffusion base frozen, SR lowered joint attribute error and FID on average relative to both uncorrected baselines (Appendix G.5, Table 45).
  3. Density ratio estimation (7uEW, jttG). We compared linear and kernel logistic regression, uLSIF, and RuLSIF, tuning each by source validation alone (Appendix G). Rankings varied with the reference and weighting rule. Within SR, direct use of bounded relative ratios gave target errors comparable to or higher than those of the original soft logistic fit. A wider regularization search reduced generated joint TV for soft logistic SR in all five development blocks. At the original guidance strength, the fixed threshold gave the lowest joint TV among the tested SR variants. Theorem 2 applies to nonnegative ratio estimates with finite, positive mean under our assumptions.
  4. TIW and LFIW (k1w6). Both methods use representative target samples. SR estimates the positive target by reweighting unlabeled data using labeled positives and selection records. Combining either method with the estimated target remains open.

Remaining concern about CelebA64 image quality (jttG). Following jttG's suggestions, we examined both ratio estimation and diffusion sampling. The reviewer acknowledged the improvements and maintained a rating of Weak Rejection, citing image quality. Thresholded SR lowered FID relative to the unguided base on all ten fresh blocks (23.93 versus 25.66 on average, Table 45). On the development blocks, stronger guidance brought the generated attribute mix closer to the target. FID at that setting remained about 7 points higher than for real images resampled with the same weights (Appendix G.4). Figure 10 shows uncurated samples from the main comparison.

Best regards,

The authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.