{"id":"d426e3f7-3c98-4e73-b796-e787ced7e485","arxiv_id":"1908.06603","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"TL-LLP is a transfer-learning extension of SVR-based label-proportions classification that models each input as lying in a bounded noise region and is claimed to be more accurate and noise-robust, though the uncertainty update actually fits the noise.","lead":"This paper combines learning from label proportions, transfer learning, and bounded-noise uncertainty into one SVM-style classifier that can use source data to help train on target data. It reports accuracy gains over four single-task LLP baselines on text datasets, but the noise-handling claim is not supported by the model or the experiments.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The uncertainty update in problem (5) minimizes over Delta x, fitting noise rather than being robust to it, so the 'less sensitive to noise' claim is unsupported.","rationale":"The reader's weakest_assumption correctly identifies the single load-bearing issue. Problem (5) and Lemma 3 show that Delta x is chosen to shrink training slack, so the uncertainty model is an optimistic label-fitting mechanism, not a robust classifier. This is not merely a disagreement with a particular robustness convention: the paper's own Section 4.4 uses the update to claim noise insensitivity, and the deployed classifier has no access to Delta x. The connection between the uncertainty formulation and the empirical noise-sensitivity result is therefore broken. My read does not change the REJECT verdict; the transfer-LLP combination might be salvageable, but as written the central robustness claim is unsupported. I also note the missing target-only and transfer-LLP baselines, but the stronger and more specific problem is the min-min uncertainty update.","tokens_in":18393,"tokens_out":3796,"duration_ms":41758,"concrete_test":"Run the Section 4.4 noise experiment on the same nine datasets in two configurations: (A) full Algorithm 1 with Lemma 3 Delta updates; (B) the identical optimization with Delta x fixed to zero throughout, i.e., solve problem (6) once. Compare test accuracy on noisy target data as the noise percentage goes from 0% to 32%. If configuration B matches or exceeds A, the Lemma 3 update is not the source of noise robustness and the 'less sensitive to noise' claim fails. A complementary check: record the average training slack immediately before and after the Delta update; Lemma 3 implies it strictly decreases, confirming that the update is an optimistic data-pushing operation rather than a worst-case robust one.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central noise-robustness claim depends on problem (5), where Delta x is a decision variable with ||Delta x|| <= delta_i. For any fixed hyperplane, the slack constraints are relaxed by choosing Delta x that moves each training point toward the margin. Lemma 3 makes this explicit: when f_t(x) - y_t > epsilon, Delta x = -delta u/||u||, the direction decreasing f_t(x), and symmetrically for the opposite violation. Thus the joint optimization is a min-min problem over (w, b, xi) and over Delta x, selecting the most favorable perturbation for each training example. A robust treatment of bounded input uncertainty would instead require constraints to hold for every Delta x in the uncertainty ball, i.e., a min-max formulation. At test time, the deployed classifier f2(x) = (w0 + v2)^T x + b2 has no access to Delta x (Algorithm 1 returns only f1 and f2; Section 3.2). Consequently, the training procedure fits the training noise rather than learning invariance to it. The accuracy gains in Table 2 over single-task baselines are expected because TL-LLP uses additional source-task data, and Section 4.4 only shows graceful degradation, not that the Delta x mechanism provides robustness. The noise-robustness claim is therefore not supported by the formulation as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes TL-LLP, a transfer learning method for learning with label proportions where both source and target data are subject to bounded additive uncertainty. The model introduces a shared parameter w0 and task-specific increments v1 and v2, and it represents uncertainty by allowing each input x to be shifted by a vector Δx with ||Δx|| ≤ δ_i. An alternating algorithm is proposed: for fixed Δx, a dual QP is solved; for fixed hyperplanes, Δx is updated via a closed-form expression. Experiments on nine text classification tasks compare TL-LLP with four LLP baselines and include a noise-sensitivity study.","tokens_in":18615,"tokens_out":5462,"duration_ms":53257,"significance":"If the central claims were valid, the paper would make a useful contribution by combining transfer learning with LLP under input uncertainty. The dual derivation for fixed Δx is standard, and the experimental study is fairly extensive. However, the central robustness mechanism is a minimization over the noise variables, which reduces training slack rather than providing worst-case robustness; this is a fundamental mismatch with the paper's stated claim. In addition, the experiments do not isolate the uncertainty-handling benefit from the use of additional source data. The paper includes a complete derivation of the dual and a reproducible experimental protocol, but the load-bearing noise-robustness claim is not supported by the formulation as written.","major_comments":[{"comment":"The uncertainty model is not robust. For a fixed hyperplane, the slack constraints are relaxed by choosing Δx that moves each training point toward the margin, as made explicit by Lemma 3, Eq. (11). The joint optimization over (w,b,ξ,Δx) is a min-min problem that selects the most favorable perturbation for each training example. A robust treatment of bounded input uncertainty would require the constraints to hold for every Δx with ||Δx|| ≤ δ_i, i.e., a min-max formulation. Since Algorithm 1 returns only f1 and f2 and the deployed classifier has no access to Δx at test time, the training procedure fits the training noise rather than learning invariance to it. Consequently, the 'less sensitive to noise' claim in the abstract and Section 4.4 is not supported by the formulation as written.","section":"Section 3.2, problem (5) and Lemma 3"},{"comment":"The reported accuracy gains over MeanMap, InvCal, alter-∝SVM and conv-∝SVM are confounded by the fact that TL-LLP has access to source-task data while the baselines are trained only on the target task. The higher accuracy in Table 2 and the more moderate degradation in Figure 5 could be entirely due to the additional source data. An ablation that disables the Δx update (i.e., fixes Δx = 0 in the same transfer model) is required to attribute the robustness to the uncertainty mechanism. Without such a comparison, Section 4.4 does not demonstrate that TL-LLP 'reduces the effect of noises in the data.'","section":"Section 4.3 and 4.4, Tables 2 and Figure 5"},{"comment":"The expression 'δti−ut||ut||' in Eq. (11) and in the proof is dimensionally inconsistent (δ is a scalar, ut/||ut|| is a unit vector) and should presumably read −δti·ut/||ut||. As written, the update rule is not well-defined, which affects the core alternating step in Algorithm 1 and the reproducibility of the method.","section":"Equation (11) and Section 6.3"},{"comment":"No convergence guarantee is given for the alternating procedure. The update of Δx uses a first-order Taylor expansion of the kernel and fixes the Lagrange multipliers, so it is not shown that the objective in problem (5) decreases monotonically or that the stopping criterion reflects convergence to a stationary point. This leaves the algorithm's behavior unspecified and is a load-bearing gap for a method whose main contribution is the iterative framework.","section":"Algorithm 1 and Section 6.3"}],"minor_comments":[{"comment":"The sentence 'For the instance x1i in target task, if f1(x1i) = w^T_2·x1i + b2 > 0' appears to use the wrong function name; it should presumably be f2(x1i) = (w0+v2)^T x1i + b2.","section":"Section 3.2, after Algorithm 1"},{"comment":"The index notation in the definitions of u1 and u2 is inconsistent, e.g., 'xmn' and 'BTm' appear in the main text while the appendix uses different target-task indices. This makes the formulas hard to follow and should be unified.","section":"Lemma 3, definitions of u1 and u2"},{"comment":"The abstract claims the method is 'less sensitive to noise,' but the experiments only report accuracy at different noise percentages; a direct measure of sensitivity (e.g., the slope of accuracy versus noise level) would be more directly aligned with the claim.","section":"Section 4.4"},{"comment":"There are several typos, including 'bond score δi' (should be 'bound score') and 'standard derivation' (should be 'standard deviation'). References [2] and [3] are the same work in different venues and should be cross-referenced.","section":"Global"}],"recommendation":"reject","confidential_remarks":"The paper's core robustness claim is not supported by the formulation, which minimizes over the noise variables rather than adopting a worst-case robust treatment. The transfer-learning component might be salvageable in a substantially revised manuscript, but the current version does not support the stated contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Honest take: the paper is a readable, competent merge of three known ideas—SVR-based LLP, regularized multi-task learning, and bounded-norm input uncertainty. The specific combination (transfer learning for LLP where both tasks have uncertain inputs) is new in the cited literature, and the authors provide a full dual derivation, an alternating algorithm, and experiments on nine text datasets. As an engineering combination, it is mostly sound.\n\nThe trouble is the uncertainty modeling, and it is load-bearing. In problem (5) the Delta x variables are part of the optimization; Lemma 3 shows they are chosen to push misclassified points toward the margin, reducing slack. That is a min-min scheme—the optimizer picks the friendliest perturbation for each training point—rather than a robust min-max scheme where the constraints must hold for every Delta x in the uncertainty ball. At test time, f2(x) has no access to Delta x, so the model has learned to fit the noise, not to be invariant to it. The paper's own text says \"we can determine a choice of ∆xi to render xi +∆xi\" and claims this makes TL-LLP less sensitive to noise; that does not follow.\n\nThe experiments do not rescue it. All four baselines are single-task LLP methods that train only on the target data, while TL-LLP sees the source data too. The gains in Table 2 are exactly what you'd expect from the extra source task. Section 4.4 only shows graceful degradation as noise fraction increases; it doesn't isolate the Delta x mechanism. The right control would be a transfer-LLP variant without the uncertainty term, or a target-only TL-LLP, plus a genuinely robust min-max baseline.\n\nThe citation pattern is fine—heavy self-citation, but the borrowed ingredients come from the right sources (Bi and Zhang, Evgeniou-Pontil, Rueping) and no circular reasoning is used.\n\nBottom line: this is a paper for practitioners who want a worked example of transfer learning in LLP; the algebra is checkable and the algorithm is implementable. But the headline claim about robustness is unsupported, and the experimental comparison is mismatched. It deserves a serious referee because the flaw is substantive and instructive, and the transfer-LLP idea might be salvageable with a correct robust formulation. I would not cite it as evidence of noise robustness.","headline":"Readable transfer-LLP package whose stated noise robustness is undermined by a min-min formulation that fits training noise rather than defending against it.","tokens_in":19184,"tokens_out":3513,"would_cite":false,"duration_ms":34497,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proposes TL-LLP, a transfer-learning method for learning with label proportions that models each input's uncertainty as a bounded perturbation and, on nine text datasets, reports higher accuracy and lower noise sensitivity than…","keywords":["learning with label proportions","transfer learning","uncertain data","support vector regression","noise robustness","alternating optimization","bag-level supervision"],"falsifier":"The paper's noise experiments corrupt the training data (Figure 5). The decisive check is to corrupt only the target test set after training, with training data clean, and compare TL-LLP with a version whose $\\Delta x$ machinery is disabled; if the accuracy gap vanishes or reverses, the reported noise insensitivity is an artifact of training-time perturbation rather than genuine robustness to input uncertainty.","tokens_in":18106,"feed_emoji":"📊","tokens_out":9537,"duration_ms":84215,"temperature":0.7,"pith_summary":"Learning with label proportions (LLP) supplies only bag-level statistics—the fraction of positive instances in each bag—not instance labels. This paper claims that when both the source and target tasks contain uncertain or noisy data, a classifier can be improved by transferring knowledge from the source task while explicitly modeling each input as lying within a bounded 'reachability area' around the observed point. The proposed TL-LLP method decomposes the source and target weight vectors as $w_1 = w_0 + v_1$ and $w_2 = w_0 + v_2$, so a shared parameter $w_0$ carries the transferable structure, and solves an alternating SVR-style optimization that also chooses perturbation vectors $\\Delta x_i$ inside their bounds. On nine text dataset pairs, the paper reports that TL-LLP achieves higher mean accuracy and smaller standard deviations than MeanMap, InvCal, alter-∝SVM, and conv-∝SVM, and that its accuracy degrades more slowly as up to 32% of training data is corrupted by Gaussian noise.","feed_headline":"Transfer learning improves label-proportion classification under noise","feed_subtitle":"A shared decision boundary plus bounded noise modeling beats four existing LLP methods on nine text datasets.","key_machinery":"The load-bearing objects are the weight decomposition $w_t = w_0 + v_t$ and the reachability area of each instance. The decomposition is what makes transfer possible: $w_0$ is the common direction learned from both tasks, and $v_1, v_2$ are small task-specific corrections, so the target classifier $f_2$ inherits structure from the source. The reachability area is the set of possible true inputs $\\{x_i + \\Delta x_i : \\|\\Delta x_i\\| \\le \\delta_i\\}$; by making $\\Delta x_i$ a variable in problem (5), the model lets the optimizer choose perturbations that reduce the training slack, which is the mechanism behind the claimed noise insensitivity. Lemmas 2 and 3 are the workhorses: Lemma 2 dualizes the QP after fixing $\\Delta x$, giving closed-form expressions for $w_0, v_1, v_2$; Lemma 3 solves the $\\Delta x$-update in closed form using the gradient direction of the kernel, effectively pushing each noisy point back inside the $\\varepsilon$-tube at cost $O(|D_1|+|D_2|)$ per iteration.","core_discovery":"The paper's central claim is that label-proportion learning can be made transferable and noise-tolerant at the same time by solving a single objective over two coupled SVR problems. For source task $T_1$ and target task $T_2$, the classifier planes are $f_1(x) = (w_0 + v_1)^T x + b_1$ and $f_2(x) = (w_0 + v_2)^T x + b_2$; the shared vector $w_0$ is what carries knowledge from source to target, while $v_1$ and $v_2$ absorb task-specific deviations. To handle uncertainty, each instance $x_i$ is assumed to be observed with additive noise $\\Delta x_i$ satisfying $\\|\\Delta x_i\\| \\le \\delta_i$, so the true input lies in a reachability area; the optimization (problem (5)) uses the perturbed inputs in the bag-average constraints and includes the $\\Delta x_i$ as decision variables. The alternating Algorithm 1 fixes the noise vectors and solves a convex QP dual (Lemma 2) to update all hyperplane parameters, then fixes the hyperplanes and updates each $\\Delta x_i$ by the closed form of Lemma 3, which moves a perturbed point toward the correct side of the classifier whenever it falls outside the $\\varepsilon$-insensitive zone. The paper takes the resulting $f_2$ as the target classifier and reports, in Table 2 and Figure 5, that it beats the four baselines in accuracy and in robustness to added Gaussian noise over nine transfer dataset pairs.","pith_inferences":["Extension: Lemma 3 chooses $\\Delta x_i$ that shrink training slack, an optimistic perturbation; if noise is adversarial or present only at test time, the robustness claim may not carry over, since at test time the model evaluates $f_2(x)$ on the observed, unperturbed point.","Extension: The weight decomposition assumes the source and target tasks share a common direction $w_0$; on unrelated task pairs the transferred component could hurt, a regime the paper does not test.","Extension: A decisive comparison would be TL-LLP with the $\\Delta x$ terms removed; this would isolate whether the reported gains come from transfer, from uncertainty modeling, or from their interaction.","Extension: Because the derivation uses kernel evaluations and a Taylor expansion in Lemma 3, the same transfer-and-perturb strategy extends to nonlinear kernels, although the experiments use a linear kernel."],"forward_implications":["If TL-LLP works as reported, LLP practitioners with noisy text data can improve target-task accuracy by adding a related source task and estimating a per-sample noise bound $\\delta_i$, without needing instance labels in either task.","The shared $w_0$ means the target classifier is not trained from scratch: the source task contributes the common decision direction, which is why the method can outperform single-task LLP baselines that see only the target bags.","The closed-form $\\Delta x$ update in Lemma 3 keeps the added computational cost linear per iteration, so total training time stays at $m \\cdot O((|D_1|+|D_2|)^2)$, dominated by the QP solve.","The reported lower standard deviations across bag sizes suggest that transfer plus uncertainty modeling stabilizes predictions, not just raises mean accuracy."],"supporting_citations":[{"why":"Defines the MeanMap baseline that estimates class means from bag proportions and is compared against in all experiments.","marker":"[3]"},{"why":"Introduces the proportion-SVM formulation and the alter-∝SVM and conv-∝SVM baselines that TL-LLP must beat.","marker":"[9]"},{"why":"Proposes SVM estimation from group probabilities (InvCal), the inverse-calibration baseline used for comparison.","marker":"[22]"},{"why":"Provides the Platt scaling function whose inverse maps bag proportions to the SVR targets y in equation (1).","marker":"[43]"},{"why":"Supplies the total-least-squares alternating optimization idea behind Algorithm 1's iterative scheme.","marker":"[44]"},{"why":"Establishes the bounded input-uncertainty model with a norm constraint on additive noise, the basis of the reachability area.","marker":"[45]"},{"why":"Provides the regularized multi-task weight decomposition that motivates writing w_t = w_0 + v_t.","marker":"[16]"},{"why":"Supplies the transfer text-classification setting and the 20 Newsgroups-derived dataset construction used in experiments.","marker":"[49]"}],"fun_headline_variants":["Transfer learning tames noisy label-proportion data","Shared boundary beats noise in label-proportion learning","Coupled tasks boost label-proportion learning under uncertainty","Noise-aware transfer learning for label proportions","Transfer learning makes label-proportion learning noise-proof"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that letting the learner move each noisy input to a nearby point that reduces its training error is a faithful model of uncertainty; if real noise does not behave that way, the claimed robustness need not transfer to test time.","fun_headline_variants_meta":{"raw":{"variants":["Transfer learning tames noisy label-proportion data","Shared boundary beats noise in label-proportion learning","Coupled tasks boost label-proportion learning under uncertainty","Noise-aware transfer learning for label proportions","Transfer learning makes label-proportion learning noise-proof"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001326,"raw_usage":{"total_tokens":5425,"prompt_tokens":1002,"completion_tokens":4423,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":618,"completion_tokens_details":{"reasoning_tokens":4352}},"tokens_in":618,"tokens_out":4423,"duration_ms":33965,"temperature":1.0,"reasoning_tokens":4352,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:39:48.305054+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"The paper's noise experiments corrupt the training data (Figure 5). The decisive check is to corrupt only the target test set after training, with training data clean, and compare TL-LLP with a version whose $\\Delta x$ machinery is disabled; if the accuracy gap vanishes or reverses, the reported noise insensitivity is an artifact of training-time perturbation rather than genuine robustness to input uncertainty.","supporting_citations":[{"cited_title":"Estimating labels from label proportions","cited_arxiv_id":null,"evidence_quote":"Defines the MeanMap baseline that estimates class means from bag proportions and is compared against in all experiments."},{"cited_title":"Svm classifier estimation from group probabilities","cited_arxiv_id":null,"evidence_quote":"Proposes SVM estimation from group probabilities (InvCal), the inverse-calibration baseline used for comparison."},{"cited_title":"Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods","cited_arxiv_id":null,"evidence_quote":"Provides the Platt scaling function whose inverse maps bag proportions to the SVR targets y in equation (1)."},{"cited_title":"The total least squares problem: Computational aspects and analysis","cited_arxiv_id":null,"evidence_quote":"Supplies the total-least-squares alternating optimization idea behind Algorithm 1's iterative scheme."},{"cited_title":"Support vector machines with input data uncertainty","cited_arxiv_id":null,"evidence_quote":"Establishes the bounded input-uncertainty model with a norm constraint on additive noise, the basis of the reachability area."},{"cited_title":"Regularized multi–task learning","cited_arxiv_id":null,"evidence_quote":"Provides the regularized multi-task weight decomposition that motivates writing w_t = w_0 + v_t."},{"cited_title":"Transferring naive bayes classifiers for text classification","cited_arxiv_id":null,"evidence_quote":"Supplies the transfer text-classification setting and the 20 Newsgroups-derived dataset construction used in experiments."}],"review_version":1}