acceptodds
Under review as a conference paper at ICLR 2027

Learning Robust Recommenders from Noisy Implicit Feedback via GMM-Weighted Bayesian Transition Matrix

Abstract

Label noise is a central challenge in learning from implicit feedback for recommendation. Conventional approaches discard noisy examples for robustness, but this sacrifices data efficiency. Unlike filtering approaches, Bayes-label transition matrix (BLTM) based methods keep all data, but their transition matrix estimates are skewed in practice. To reduce this skew, we introduce GMM-weighted Bayes-label Transition Matrix (RGBT), which augments BLTM with GMM-based instance weights. A GMM assigns each instance a reliability score, and these scores calibrate the BLTM to reduce bias. We show theoretically that RGBT retains all samples, yields consistent estimates, and provably reduces variance compared to CLTM, generalizing to any method with deterministic labels and non-negative weights. Experiments on real and synthetic datasets show that RGBT handles noisy samples more effectively than sample-selection methods, and calibrates the transition matrix more accurately than existing approaches.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Noisy Feedback, Gaussian Mixture Model, Bayes-label Transition Matrix, Recommender Systems

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_JtFJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes RGBT for learning recommender models from noisy implicit feedback. It combines Bayes-label transition matrix (BLTM) with GMM-based reliability weighting, where the GMM uses prediction scores and user-item co-occurrence features to estimate sample reliability. Experiments on three datasets and six backbones show improved recommendation performance in many settings, and lower average transition matrix estimation error under synthetic noise. The paper also claims consistency and variance reduction guarantees.

Critical strengths
  1. The idea of using reliability estimation together with transition matrix correction is reasonable. Keeping all noisy interactions in the main training objective can be useful, especially when reliable interactions are limited.

  2. The empirical results are quite good in some settings, particularly LightGCN and XSimGCL. Also good to see the authors directly evaluate transition matrix estimation under controlled noise, rather than only reporting downstream recommendation metrics.

Critical weaknesses
  1. I have some concerns about the theoretical analysis. The Bayes optimal classifier is not necessarily preserved under instance-dependent noise. Also, training with hard pseudo-labels cannot generally recover the clean class posterior, which makes the consistency claims questionable. The variance reduction proof also seems to assume equal conditional variances for different estimators, without sufficient justification.

  2. The contribution of GMM is not very clear to me. The reliability weights are applied to pseudo-label learning, but do not directly enter the transition matrix loss. From the ablation results, removing GMM only causes small performance drops, while the improvement over original BLTM is much larger. So how much improvement actually comes from GMM calibration? The current experiments do not quite explain this.

Questions for authors
  1. How can Theorem 3 recover a non-one-hot clean posterior from hard Bayes labels? What additional assumptions make the transition matrix identifiable?

  2. Where exactly do GMM weights enter the transition network update? Can authors report matrix estimation error for w/o GMM under the same training setup?

  3. Does the variance inequality still hold with highly concentrated GMM weights? The derivation of Eq. (30) seems not valid to me.

6 comments
AuthorsOpenReview ↗

Response for Submission2305 by Reviewer JtFJ

Thank you for the valuable and insightful comments. We will revise the submission according to the comments. Our responses are as follows.

Q1: IDN breaks Bayes preservation; hard labels can't yield soft posteriors; variance proof assumes equal conditional variances.

A1: The Bayes-optimal classifier is not claimed to be preserved in general. Assumption 1 (Sec. 3.2.1) only requires Bayes-label preservation on the reliable set Xrel\mathcal{X}_{\mathrm{rel}}, and nothing is claimed off Xrel\mathcal{X}_{\mathrm{rel}}; Theorem 1 and Theorem 3 are both stated on Xrel\mathcal{X}_{\mathrm{rel}}.

Hard pseudo-labels cannot recover a non-one-hot clean posterior. Theorem 3 (Sec. 3.2.2) claims f→ηf\to\boldsymbol{\eta} only on Xrel\mathcal{X}_{\mathrm{rel}}. The new {Scope of Theorem 3} paragraph states that off Xrel\mathcal{X}_{\mathrm{rel}}, ff converges only to the one-hot distilled-label distribution π∗(x)\boldsymbol{\pi}^*(\mathbf{x}), with no claim of recovering η(x)\boldsymbol{\eta}(\mathbf{x}); the (0.7,0.3)(0.7,0.3) case is a limitation of the information content of hard labels. Theorem 4 is restricted to Xrel\mathcal{X}_{\mathrm{rel}} for the same reason and explicitly depends on Assumptions 2--4 and the basin condition.

The two conditional variances are no longer assumed equal. Their difference Δcond\Delta_{\mathrm{cond}} is kept explicitly and bounded as ∣Δcond∣≤C(K,ρmax⁡)(wmax⁡2/wmin⁡2)neff−1|\Delta_{\mathrm{cond}}|\le C(K,\rho_{\max})(w_{\max}^2/w_{\min}^2)n_{\mathrm{eff}}^{-1}, see the {Conditional decomposition} paragraph of Sec. 3.2.4 and the term-by-term derivation in Appendix A.5. Theorem 5 imposes the domination condition, which ensures that κij(x)\kappa_{ij}(\mathbf{x}) dominates ∣Δcond∣|\Delta_{\mathrm{cond}}|; when this condition fails, strict variance reduction is not claimed, and Corollary 1 degrades as neff−1/2n_{\mathrm{eff}}^{-1/2}, see the {When the condition fails} paragraph of Sec. 3.2.4, the {Conclusion} paragraph of Appendix A.5, and Appendix A.6.

Q2: GMM’s contribution is unclear: its weights bypass the transition loss, and ablations show only marginal gains.

A2: In the revised Lclass\mathcal{L}_{\mathrm{class}} (Sec. 3.1), GMM weights directly scale distilled samples' contributions to the transition network update, in addition to weighting distilled label supervision of the classifier (details and Table 2 errors are provided in A4).

Compared with BLTM, both Full and w/o GMM incorporate two designs described in Sec. 3.1 and Algorithm 1: direct supervision of the classifier using distilled labels, and alternating updates of the classifier and transition network. Full versus w/o GMM measures the additional effect of reliability weighting. Table 2 (Sec. 4.3) reports a lower average transition matrix estimation error with GMM under both noise types.

Q3: How does Theorem 3 recover soft posteriors from hard labels, and what ensures identifiability?

A3: Theorem 3 does not claim to recover a non-one-hot clean posterior from hard labels. It claims f→ηf\to\boldsymbol{\eta} only on the reliable set Xrel\mathcal{X}_{\mathrm{rel}}, where η\boldsymbol{\eta} is one-hot and hard labels carry the same information as the clean posterior; off Xrel\mathcal{X}_{\mathrm{rel}}, ff converges only to the one-hot distilled-label distribution π∗(x)\boldsymbol{\pi}^*(\mathbf{x}), and no recovery of η(x)\boldsymbol{\eta}(\mathbf{x}) is claimed, with (0.7,0.3)(0.7,0.3) being a limitation of the information content of hard labels. Theorem 3 relies on Assumption 1 (Bayes-preserving on Xrel\mathcal{X}_{\mathrm{rel}}) and Assumption 2 (uniformly reliable distillation).

Identifiability of the transition matrix is supplied by Assumption 3 (local identifiability) and Assumption 4 (Lipschitz transition matrix): Assumption 3 requires the Jacobian of ϕ↦g(x;ϕ)⊤η(x)\phi \mapsto g(\mathbf{x};\phi)^\top \boldsymbol{\eta}(\mathbf{x}) to have full column rank at ϕ∗\phi^*, giving pointwise uniqueness, and Assumption 4 requires T⋆(x)T^{\star}(\mathbf{x}) to be Lipschitz in x\mathbf{x}, controlling the cross-instance approximation error. The {Identifiability} paragraph of Sec. 2 notes that η~=T⋆⊤η\tilde{\eta}=T^{{\star}\top}\eta provides only K−1K-1 equations against K(K−1)K(K-1) degrees of freedom, so identification does not follow from this relation alone; this is consistent with Liu et al. (ICML 2023), who show that instance-dependent transition matrices are generally not identifiable from a single noisy label without additional assumptions. These two assumptions supply the missing structure.

AuthorsOpenReview ↗

Response for Submission2305 by Reviewer JtFJ

Q4: How do GMM weights affect the transition update, and what is the error without GMM?

A4: In the revised Lclass\mathcal{L}_{\mathrm{class}} (Sec. 3.1), the loss is computed using classifier predictions adjusted by the estimated transition matrix, so it directly depends on the transition network parameters. Multiplying each distilled sample's loss by its GMM weight therefore scales its gradient contribution to the transition network update, while samples outside the distilled set keep unit weight.

Table 2 (Sec. 4.3) compares Full and w/o GMM, which sets all sample weights to one, under the same training setup.

Mean entrywise L1L_1 error on MovieLens, averaged over four backbones and four noise rates. Lower is better.

Noisew/o GMMFull
Symmetric1.13011.1050
Pairflip2.23932.2240

We measure the entrywise L1L_1 distance between the estimated matrix averaged over validation interactions and the known transition matrix.

Q5: Does the variance bound hold for concentrated weights, and is Eq. (30) valid?

A5: The variance inequality is not guaranteed when the GMM weights are highly concentrated. In that regime neffn_{\mathrm{eff}} shrinks and the bound on Δcond\Delta_{\mathrm{cond}} grows; once the domination condition neff≥C(K,ρmax⁡)wmax⁡2/(wmin⁡2κij(x))n_{\mathrm{eff}}\ge C(K,\rho_{\max})w_{\max}^2/(w_{\min}^2\kappa_{ij}(\mathbf{x})) fails, κij(x)\kappa_{ij}(\mathbf{x}) need not dominate ∣Δcond∣|\Delta_{\mathrm{cond}}|. Theorem 5 therefore claims strict variance reduction only under the domination condition; when it fails, strict variance reduction is not claimed, and Corollary 1 degrades as neff−1/2n_{\mathrm{eff}}^{-1/2} (Sec. 3.2.4, {When the condition fails}; Appendix A.5, {Conclusion}; Appendix A.6).

Eq. (30) is rewritten: the two conditional variances are no longer assumed equal, and their difference Δcond\Delta_{\mathrm{cond}} is kept explicitly and bounded as ∣Δcond∣≤C(K,ρmax⁡)(wmax⁡2/wmin⁡2)neff−1|\Delta_{\mathrm{cond}}|\le C(K,\rho_{\max})(w_{\max}^2/w_{\min}^2)n_{\mathrm{eff}}^{-1}. The derivation is in the {Conditional decomposition} paragraph of Sec. 3.2.4 and in Appendix A.5, which states explicitly that the two conditional variances are not assumed equal and that their difference is bounded.

Please refer to the latest version of the PDF for all specific modifications.

Liu Y, Cheng H, Zhang K. Identifiability of label noise transition matrix[C]//International Conference on Machine Learning. PMLR, 2023: 21475-21496.

Thanks for the detailed response and for fixing the theoretical issues.

I appreciate that the authors revised the theorems carefully, restricting Theorems 1, 3, and 4 to XrelX_{rel}, adding the identifiability assumptions, and correcting Eq. (30) with the explicit Δcond\Delta_{cond} bound and domination condition. These changes make the math much more rigorous.

However, after these revisions, the core contribution feels noticeably narrower. The consistency guarantees now only hold on Xrel{X}_{rel} where the clean posterior is already one-hot, and strict variance reduction requires a domination condition that can easily break when GMM weights become concentrated. More importantly, the new ablation numbers in Table 2 show that GMM weighting only brings a tiny reduction in matrix estimation error over w/o GMM (e.g., 2.2393 to 2.2240 under Pairflip), and the component analysis (in response to dQxs) shows that under Pairflip noise, the selected GMM component actually has a higher label error rate (23.74% vs. 16.44%) and lowers the clean label proportion (79.97% to 77.64%). Since most of the empirical gain seems to come from direct distilled-label supervision and alternating updates rather than GMM calibration, I will keep my rating at 2.

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

This paper proposes a novel recommender model which combines the merits of BLTM and GMM. It is data-efficient by using all the data, gives consistent estimates, and shows lower variance than CLTM. Theorems are proven to show its theoretical edge. Experiments are conducted to show validity of the model.

Critical strengths
  1. This is a novel structure of combining BLTM and GMM. I like this idea of joint learning of both the calibration loss and the classification loss.

  2. The experiment part is thorough, by including performance comparison, robustness of calibration, hyper-parameter sensitivity, and ablation. The experiments do establish the validity of the algorithm, I believe.

Critical weaknesses

Readability is a big issue of this paper:

(1) Abbreviations are used before introducing the full name, for example in the abstract.

(2) Many quantities are defined multiple times. For example, equation (1) is at least defined 3 times.

(3) Quantities used in theorems are not even defined. For example, I have no idea what is P_D or g^* in Theorem 1.

People might get around these issues by reading the paper many times, but I believe this is not good practice and the paper needs to improve its structure and writing.

Questions for authors

Q1 How should I interpret the conclusion of Theorem 2? Why does η>1/2\eta > 1/2 indicate ``better than random’’?

Q2 Why use GMM? Is there any specific observation or motivation for this choice?

Q3 You mentioned ``extending beyond GMM-based calibration to any method with deterministic labels and non-negative weights’’. Can you elaborate more on this?

3 comments
AuthorsOpenReview ↗

Response for Submission2305 by Reviewer mWTz

Thank you for the valuable and insightful comments. We will revise the submission according to the comments. Our responses are as follows.

Q1: Abbreviations are used before being defined.

A1: We have revised the manuscript so that all abbreviations (RGBT, GMM, CLTM, etc.) are fully defined at first occurrence, including in the Abstract, before use.

Q2: Redundant definitions (e.g., Eq. (1) defined three times).

A2: We have removed the duplicate definitions of the distilled labels (formerly Eqs. 1, 6, and 18) in Sections 3.1.1 and 3.2.2; these are now defined once in Section 2 and referenced thereafter. Repeated notations (T∗T^*, ρmax⁡\rho_{\max}, Ddistilled\mathcal{D}_{\mathrm{distilled}}) have also been cleaned up across the Method and Theory sections.

Q3: Undefined symbols in theorems (e.g., �0��0�, �1��1�).

A3: We have added explicit definitions in Section 2: PDP_D is the clean joint distribution and g∗(x)g^*(\mathbf{x}) is the Bayes classifier. All symbols used in the Method and Theory sections are now defined before use.

Q4: Why does �2��2� imply “better than random” in Theorem 2?

A4: Theorem 2 states that when the retained label's confidence exceeds (1+ρmax⁡)/2(1+\rho_{\max})/2, its clean probability satisfies ηy∗(x)>1/2\eta_{y^*}(\mathbf{x})>1/2. Random guessing has accuracy 1/K1/K, and 1/2>1/K1/2>1/K for all K≥2K\ge 2, so the retained label is strictly better than random guessing; the 1/21/2 is an absolute majority bound independent of KK. The derivation is in the new {Why the threshold is (1+ρmax⁡)/2(1+\rho_{\max})/2} paragraph of Section 3.2.2 (Eq. 16) and in Appendix A.4. The paragraph after Theorem 2 also states that this does not imply hard labels recover a soft posterior such as (0.7,0.3)(0.7,0.3).

Q5: Why use a GMM specifically?

A5: We use a two-component GMM because distilled samples are not equally reliable. Two signals, the prediction score y^uv=pu⊤qv\hat{y}_{uv}=\mathbf{p}_u^\top\mathbf{q}_v and the co-occurrence count cuvc_{uv}, separate pairs that are strong on both signals from atypical ones, and a two-component mixture is the minimal model that captures this bimodal reliability structure without imposing a hard threshold. The GMM assigns bounded weights in (0,1)(0,1), down-weighting the atypical group while retaining all distilled samples. Since distillation already removes most noise, this weighting serves as a secondary calibration rather than the main denoising mechanism. The reference component is treated only as a reference group, and we do not claim it has lower label noise. The theoretical result applies to any bounded deterministic weighting, with GMM weighting as a natural instance.

Q6: How does it generalize beyond GMM to arbitrary non-negative weightings?

A6: This claim is now stated precisely as the domination condition in Theorem 5. The variance-reduction argument does not rely on the GMM specifically: it only requires (i) the label proxy used in the transition-matrix estimator to be deterministic given x\mathbf{x}, so that the first term in the law-of-total-variance decomposition vanishes, and (ii) the sample weights wkw_k to be bounded in [wmin⁡,wmax⁡][w_{\min},w_{\max}] with wmin⁡>0w_{\min}>0, so that neffn_{\mathrm{eff}} is well defined. Under these two properties, the difference between the CLTM and BLTM variances is κij(x)+Δcond\kappa_{ij}(\mathbf{x})+\Delta_{\mathrm{cond}}, with ∣Δcond∣≤C(K,ρmax⁡)(wmax⁡2/wmin⁡2)neff−1|\Delta_{\mathrm{cond}}|\le C(K,\rho_{\max})(w_{\max}^2/w_{\min}^2)n_{\mathrm{eff}}^{-1}, and the domination condition guarantees that κij(x)\kappa_{ij}(\mathbf{x}) dominates ∣Δcond∣|\Delta_{\mathrm{cond}}|, so the variance difference is non-negative. GMM weighting is one instantiation satisfying (i) and (ii); any deterministic label proxy with bounded positive weights satisfies the same argument. This generality is consistent with the fact that instance-dependent transition-matrix estimation itself requires additional structural assumptions for identifiability (Liu et al., ICML 2023). When the condition fails, strict variance reduction is not claimed, and Corollary 1 degrades as neff−1/2n_{\mathrm{eff}}^{-1/2}.

Please refer to the latest version of the PDF for all specific modifications.

Liu Y, Cheng H, Zhang K. Identifiability of label noise transition matrix[C]//International Conference on Machine Learning. PMLR, 2023: 21475-21496.

The latest version appears far more readable than the original one. Thanks for addressing my concerns. I thereby increase my score.

AuthorsOpenReview ↗

Response for Submission2305 by Reviewer mWTz

Thank you very much for your positive feedback and for acknowledging that the latest version is much more readable than the original. We sincerely appreciate your time and constructive comments, which have helped us substantially improve the manuscript. We are also grateful that you have decided to increase your score. We will continue to refine the paper to ensure it meets the highest standards.

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

The paper proposes RGBT for recommendation with noisy feedback. It uses GMM weights to train on reliable pseudo-labels and a transition matrix to model label noise. Experiments on three datasets and six backbones show gains in recommendation accuracy and lower average matrix estimation error.

Critical strengths

The experiments cover several datasets and backbones. The gains on MovieLens and Yelp suggest that combining reliability weights with transition correction can be useful across different recommendation models.

Critical weaknesses
  1. Theorem 3 does not follow from the stated loss. Even without noise, if the clean class probabilities are (0.7, 0.3), training on the hard Bayes label targets (1, 0), not (0.7, 0.3). Theorem 4 relies on this result. Also, matching the noisy class probabilities does not uniquely determine an instance-dependent transition matrix without further assumptions.

  2. The variance proof has a gap. Equations 28–30 cancel the conditional variance terms of two different estimators without showing that they are equal. Nonnegative weights alone do not guarantee lower variance, since uneven weights can reduce the effective sample size.

  3. The link between the training procedure and the theoretical estimator is unclear. Equation 14 applies GMM weights to the classifier loss, while Equation 27 estimates the transition matrix using weighted counts. The paper does not explain how Algorithm 1 leads to this estimator.

  4. The rule for selecting the reliable GMM component lacks support. A larger mean norm does not necessarily imply cleaner samples. Since the features include interaction counts, this rule may favor popular items. The paper needs evidence that the selected component has less label noise.

Questions for authors

Refer to the weaknesses above.

3 comments
AuthorsOpenReview ↗

Response for Submission2305 by Reviewer dQxs

Thank you for the valuable and insightful comments. We will revise the submission according to the comments. Our responses are as follows.

Q1: Theorem 3's claim is false, Theorem 4 depends on it, and identifiability is missing.

A1: Theorem 3 (Sec. 3.2.2) is now restricted to Xrel\mathcal{X}_{\mathrm{rel}} and depends on Assumption 1 (Sec. 3.2.1) and Assumption 2 (Sec. 3.2.2). A new {Scope of Theorem 3} paragraph states that off Xrel\mathcal{X}_{\mathrm{rel}}, ff converges only to π∗(x)\boldsymbol{\pi}^*(\mathbf{x}), with no claim of recovering η(x)\boldsymbol{\eta}(\mathbf{x}); the (0.7,0.3)(0.7,0.3) case is a limitation of hard labels. Theorem 4 (Sec. 3.2.3) is restricted to Xrel\mathcal{X}_{\mathrm{rel}} for the same reason and explicitly depends on Assumptions 2--4 and the basin condition.

Section 2 ({Identifiability of the transition matrix}, after Eq. 3) states that η~=T∗⊤η\tilde{\eta}=T^{*\top}\eta gives only K−1K-1 equations against K(K−1)K(K-1) degrees of freedom and that identification does not follow from Eq. 3 alone. This identifiability issue is fundamental: Liu et al. (ICML 2023) show that instance-dependent transition matrices are generally not identifiable from a single noisy label without additional assumptions. Section 3.2.3 introduces Assumption 3 (local identifiability) for pointwise uniqueness and Assumption 4 (Lipschitz continuity) for cross-instance error control. Appendix A.1 repeats the identifiability discussion; Appendix A.2.1 gives the off-Xrel\mathcal{X}_{\mathrm{rel}} argument; Appendix A.5 gives the term-by-term variance bound.

Q2: Flawed cancellation of unequal variances; uneven weights shrink sample size, voiding variance reduction.

A2: For the first issue, the two conditional variances are no longer assumed equal: their difference Δcond\Delta_{\mathrm{cond}} is kept explicitly and bounded in Eq. 17. See the new {Conditional decomposition} paragraph of Section 3.2.4 (Eqs. 15 and 17) and the {Conditional decomposition} and {Bounding Δcond\Delta_{\mathrm{cond}}} paragraphs of Appendix A.5.

For the second issue, Theorem 5 (Section 3.2.4) imposes the domination condition in Eq. 14. The new {When the condition fails} paragraph of Section 3.2.4 states that when this condition fails, strict variance reduction is not claimed and Corollary 1 degrades as neff−1/2n_{\mathrm{eff}}^{-1/2}; the derivation is in the {Conclusion} paragraph of Appendix A.5 and in Appendix A.6.

Q3: No derivation links the weighted classifier loss in Algorithm 1 to the weighted-count transition estimator.

A3: Lclass\mathcal{L}_{\mathrm{class}} is evaluated on the full set D\mathcal{D}, with weights w(x)w(\mathbf{x}) on the distilled set Ddistilled\mathcal{D}_{\mathrm{distilled}} and weight 11 on the non-distilled set D∖Ddistilled\mathcal{D}\setminus\mathcal{D}_{\mathrm{distilled}}. The effective counts in the gradient updates of Algorithm 1 are therefore ck=wkc_k=w_k for k∈Ddistilledk\in\mathcal{D}_{\mathrm{distilled}} and ck=1c_k=1 for k∈D∖Ddistilledk\in\mathcal{D}\setminus\mathcal{D}_{\mathrm{distilled}}, matching Eq. 6. Taking the first-order condition of Lclass\mathcal{L}_{\mathrm{class}} with respect to T^ij∗\hat{T}^*_{ij} and enforcing row normalization yields exactly Eq. 6, so Algorithm 1 and the theoretical estimator share the same stationary point. The derivation is in the new {From Algorithm 1 to Eq. 6} paragraph of Section 3.1 (with the Lagrangian in Eq. 13) and in Appendix A.2.2 (Eqs. 18--19 and the {Connection to Algorithm 1} paragraph).

Q4: GMM mean-norm selection is unjustified, popularity-biased, and lacks noise-rate evidence.

A4: The GMM takes the prediction score y^uv\hat{y}_{uv} and the cooccurrence count cuvc_{uv} as inputs, and both features are standardized before fitting. We select the component whose mean vector has the larger norm. To evaluate label correctness directly, we compare distilled labels on MovieLens 100K (NeuMF, 20% noise) against the labels before noise injection.

Distilled label error rate. Lower is better.

NoiseSelected componentOther component
Symmetric11.9820.93
Pairflip23.7416.44

Correct label proportion. Higher is better.

NoiseUniform weightsGMM weights
Symmetric83.4986.35
Pairflip79.9777.64

Under Symmetric noise, GMM weighting improves the weighted proportion of correct labels by 1.31 to 3.22 percentage points within each of the three item popularity groups. We will include the complete component error results, popularity group breakdown, and metric definitions in the appendix.

Please refer to the latest version of the PDF for all specific modifications.

Liu Y, Cheng H, Zhang K. Identifiability of label noise transition matrix[C]//International Conference on Machine Learning. PMLR, 2023: 21475-21496.

The author addressed my problem in the rebuttal, and I will improve my score

AuthorsOpenReview ↗

Response for Submission2305 by Reviewer dQxs

Thank you very much for your positive feedback and for acknowledging that we addressed your concern in the rebuttal. We sincerely appreciate your time and consideration, and we are grateful for your willingness to improve the score. We will continue to refine the paper accordingly.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Responses to Reviewers

We thank all reviewers for their careful reading. The revision addresses the main concerns as follows.

Reviewer dQxs.

Q1 (Theorem 3, Theorem 4, identifiability). Theorem 3 and Theorem 4 are now restricted to Xrel\mathcal{X}_{\mathrm{rel}}; a new {Scope of Theorem 3} paragraph states that off Xrel\mathcal{X}_{\mathrm{rel}} hard labels cannot encode a non-one-hot posterior, and no recovery of η(x)\eta(\mathbf{x}) is claimed. Section 2 adds an {Identifiability} paragraph (K−1K-1 equations vs.\ K(K−1)K(K-1) degrees of freedom, Liu et al., ICML 2023), and Assumptions 3--4 supply pointwise uniqueness and cross-instance error control.

Q2 (variance cancellation). The two conditional variances are no longer assumed equal; Δcond\Delta_{\mathrm{cond}} is bounded explicitly. Theorem 5 imposes the domination condition; when it fails, strict variance reduction is not claimed, and Corollary 1 degrades as neff−1/2n_{\mathrm{eff}}^{-1/2}.

Q3 (Algorithm 1 vs.\ estimator). Lclass\mathcal{L}_{\mathrm{class}} uses w(x)w(\mathbf{x}) on Ddistilled\mathcal{D}_{\mathrm{distilled}} and 11 elsewhere, so ck=wkc_k=w_k on Ddistilled\mathcal{D}_{\mathrm{distilled}} and ck=1c_k=1 outside. The first-order condition with row normalization yields Eq. 6, so Algorithm 1 and the estimator share the same stationary point.

Q4 (GMM component selection). A component-level analysis on MovieLens 100K shows the selected component is cleaner under Symmetric noise (11.98%11.98\% vs.\ 20.93%20.93\%), but not under Pairflip (23.74%23.74\% vs.\ 16.44%16.44\%); we therefore do not claim it has lower label noise in general. Full results will be in the appendix.

Reviewer mWTz.

Q1--Q3 (abbreviations, redundant definitions, undefined symbols). All abbreviations are defined at first occurrence, including the Abstract; distilled labels are defined once in Sec. 2; PDP_D and g∗(x)g^*(\mathbf{x}) are now explicitly defined.

Q4 (Theorem 2, >1/2>1/2). The retained label satisfies ηy∗(x)>1/2\eta_{y^*}(\mathbf{x})>1/2; since 1/2>1/K1/2>1/K for K≥2K\ge 2, it is better than random guessing. This does not imply hard labels recover a soft posterior such as (0.7,0.3)(0.7,0.3).

Q5 (why GMM). A two-component GMM is the minimal model for the bimodal reliability structure, with bounded weights in (0,1)(0,1); it is a secondary calibration, not the main denoising mechanism, and the reference component is only a reference group.

Q6 (generalization beyond GMM). The variance-reduction argument only requires a deterministic label proxy and bounded positive weights; GMM weighting is one instance, and any such weighting satisfies the same bound.

Reviewer JtFJ.

Q1 (Bayes preservation, hard labels, variance). Assumption 1 only requires Bayes preservation on Xrel\mathcal{X}_{\mathrm{rel}}; Theorem 1 and Theorem 3 are stated on Xrel\mathcal{X}_{\mathrm{rel}} only; the two conditional variances are not assumed equal, and Theorem 5 imposes the domination condition with the same failure mode as above.

Q2 (GMM contribution). GMM weights enter both Lcalibrated\mathcal{L}_{\mathrm{calibrated}} and Lclass\mathcal{L}_{\mathrm{class}}, so they scale distilled samples' contribution to the transition network update. Full vs.\ w/o GMM isolates this effect; Table 2 reports a lower average error with GMM under both noise types, though the gain is modest and not uniform.

Q3 (soft posteriors and identifiability). Same as dQxs Q1: Theorem 3 is restricted to Xrel\mathcal{X}_{\mathrm{rel}}, and Assumptions 3--4 supply identifiability.

Q4 (GMM weights in the transition update). In Lclass\mathcal{L}_{\mathrm{class}}, the loss depends on T^∗(x;θ)f(x;w)\hat{T}^*(\mathbf{x};\theta)f(\mathbf{x};\mathbf{w}); GMM weights scale distilled samples' gradient contribution to the transition network, while non-distilled samples keep unit weight. Table 2: Symmetric 1.1301→1.10501.1301\to1.1050, Pairflip 2.2393→2.22402.2393\to2.2240.

Q5 (concentrated weights and Eq. 30). The variance inequality is not guaranteed under highly concentrated weights; Theorem 5 claims strict variance reduction only under the domination condition, and Corollary 1 degrades as neff−1/2n_{\mathrm{eff}}^{-1/2}. Eq. (30) is rewritten with Δcond\Delta_{\mathrm{cond}} kept explicitly.

All specific changes are marked in blue in the revised PDF.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.