{"id":"d1c9426d-ec93-4f22-8c86-9c508a37af21","arxiv_id":"2507.20191","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"IS2C replaces importance weighting with synthetic same-class mixed samples whose label proportions match the target, plus class-conditional alignment, and claims provable control of target error.","lead":"In partial domain adaptation, the labeled training data has extra categories that never appear in the unlabeled test data. This paper counters that mismatch by generating new training examples whose category mix matches the test set, and it reports mixed benchmark results against a strong earlier method.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 1 is internally inconsistent: a probabilistic output with all components convex must be affine, yet for affine classifiers the mixture risk need not be no larger; the sampling-over-weighting advantage is unsupported.","rationale":"The central claim is not the unverified Theorem 1 by itself but the comparison it needs: IS2C is claimed to be at least as good as importance weighting because epsilon_c <= epsilon_c'. Proposition 1 is the only theorem that provides that comparison. Scrutinizing Proposition 1 shows it does not survive contact with the paper's own definitions: a K-dimensional probabilistic output whose components are all convex must be affine, and for affine probabilistic classifiers the 0-1 risk can increase under within-class mixing. The explicit U[0.2,1]/U[0.8,1] example gives epsilon_c=0.071875 > epsilon_c'=0.0625, so the proposition is false as stated, not merely hard to prove. This is stronger than the Reader's objection: the issue is not only that real softmax networks violate convexity, but that even in the protected convex regime the asserted risk comparison fails. The missing supplementary proofs and the mixed benchmark tables (MOT outperforms IS2C on three of four datasets) reinforce rejection, but the internal inconsistency is decisive on its own. I therefore keep the Reader's REJECT verdict, while crediting the paper for the clearly described algorithm, the O(n^2) ETIC variant, and the extensive empirical study; those may support the method as a heuristic, but not the stated theoretical guarantees.","tokens_in":22560,"tokens_out":19049,"duration_ms":182905,"concrete_test":"Implement the binary counterexample: X in [0,1], P(Y=1)=0.1, P(Y=2)=0.9, P(X|Y=1)=U[0.2,1], P(X|Y=2)=U[0.8,1], and classifier h.g(x)=(1-x,x). Compute the 0-1 risks epsilon_c' (theta=0) and epsilon_c (theta=0.5, with X=(X1+X2)/2). The exact values are epsilon_c'=0.0625 and epsilon_c=0.071875, so if the computation reproduces epsilon_c > epsilon_c', Proposition 1 is false as stated. Independently, locate the proof of Proposition 1 in the supplementary material and check the step where convexity of the output components is used to bound the non-convex 0-1 error.","verdict_should_be":"REJECT","load_bearing_attack":"The load-bearing step is Proposition 1 (Sec. III-B), which asserts epsilon_c(h.g) <= epsilon_c'(h.g) whenever every component of h.g is a convex function of the input, with equality for linear models. This proposition is the only result that tells the reader sampling is at least as good as importance weighting; Theorem 1 is just an upper bound on |epsilon_t - epsilon_c| and contains no comparison with the weighted-source risk. The proposition is incompatible with the paper's own model definition. Section III defines h.g as a K-dimensional probabilistic vector, so the components f_i = (h.g)_i are nonnegative and sum to 1. If every f_i is convex, then for any x1, x2 and lambda, 1 = sum_i f_i(lambda x1 + (1-lambda)x2) <= lambda sum_i f_i(x1) + (1-lambda) sum_i f_i(x2) = 1. Every Jensen inequality is therefore tight, so all f_i are affine; the premise never yields strict convexity. More seriously, the asserted inequality is false for affine probabilistic classifiers under the paper's 0-1 risk. Example: K=2, X in [0,1], P(Y=1)=0.1, P(Y=2)=0.9, class conditionals U[0.2,1] and U[0.8,1], and h.g(x) = (1-x, x). For theta=0 or 1, epsilon_c' = 0.1 * P(U[0.2,1] > 0.5) = 0.0625. For theta=0.5, the mixture of two independent U[0.2,1] variables has average CDF given by the Irwin-Hall distribution; P((U1+U2)/2 > 0.5) = 0.71875, so epsilon_c = 0.071875 > epsilon_c'. Thus Proposition 1 cannot be proved as stated. The authors' synthetic validation uses only a scalar convex output and does not test the sum-to-one K-dimensional probabilistic structure assumed in the paper.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes IS2C for partial domain adaptation. Instead of reweighting source samples, IS2C constructs a sampling domain whose label distribution matches the estimated target label distribution and whose class-conditional features are pairwise convex combinations of source features from the same class. The paper gives Theorem 1, an upper bound on |epsilon_t(h.g)-epsilon_c(h.g)| in terms of label-shift, conditional-shift, and mixture terms, and Proposition 1, which asserts that the risk on the sampling domain is no larger than the risk on the original or reweighted source under a convexity condition. An OT-based conditional alignment loss (ETIC) with a claimed O(n^2) per-iteration cost is proposed, and experiments are reported on Office-31, Office-Home, VisDA-2017, and Image-CLEF.","tokens_in":23051,"tokens_out":10760,"duration_ms":106257,"significance":"If valid, the theory would provide a principled sampling alternative to importance weighting for PDA, with a transparent decomposition of the target-versus-sampling risk gap, and the ETIC speedup would be practically useful. The paper also contains a clear ablation study and sensitivity analysis. However, the main comparison result (Proposition 1) is either vacuous under the paper's own definitions or false under the standard 0-1 argmax risk, and the experimental results do not consistently support superiority over the MOT baseline. The central claim that sampling is theoretically preferable to reweighting is therefore not established.","major_comments":[{"comment":"Under the paper's own definition in Section III, h.g is a K-dimensional probabilistic vector, so each component f_i = (h.g)_i is nonnegative and sum_i f_i = 1. If every f_i is convex, then the constant function 1 = sum_i f_i is convex, and Jensen's inequality must be tight for the sum, forcing each f_i to be affine. Hence the premise of Proposition 1 is compatible only with affine models, for which the risk is unchanged by the mixture operation under the paper's random-prediction risk. The claimed 'strictly smaller risk' for non-linear models does not follow from the stated assumption, so the proposition cannot justify preferring sampling over importance weighting.","section":"Section III-B, Proposition 1"},{"comment":"If the error is instead interpreted as the conventional 0-1 argmax error, which is the interpretation used in the experimental visualizations and decision-boundary figures, the asserted inequality is false. For example, take K=2, X in [0,1], P(Y=1)=0.1, P(Y=2)=0.9, X|Y=1 ~ U[0.2,1], X|Y=2 ~ U[0.8,1], and h.g(x)=(1-x,x). For theta=0 or 1, the risk is 0.1 * P(U[0.2,1] > 0.5) = 0.0625. For theta=0.5, X|Y=1 is the average of two independent U[0.2,1] variables and P(X > 0.5) = 0.71875, giving epsilon_c = 0.071875 > epsilon_c'. Thus the proposition cannot be correct under that risk.","section":"Section III-B, Proposition 1 (0-1 argmax interpretation)"},{"comment":"The proofs of Theorem 1 and Proposition 1 are deferred to a supplementary file that is not included in the arXiv submission. Since Proposition 1 is central to the paper's claim that sampling is at least as good as reweighting, and since that proposition appears to be unsound, the missing proofs are not merely a presentation issue; the derivation of the main theoretical guarantee cannot be checked.","section":"Section III-B, proofs"},{"comment":"The claimed empirical superiority is contradicted by the paper's own numbers. On Office-Home, MOT achieves mean accuracy 80.6% while IS2C achieves 79.2%; on VisDA-2017, MOT achieves 92.4% while IS2C achieves 89.3%; on Image-CLEF, MOT achieves 93.6% while IS2C achieves 93.2%. The sentence immediately after Table I stating that IS2C outperforms comparison methods by at least about 1.0% does not hold against MOT, so the empirical case for IS2C over the strongest baseline is not established.","section":"Section V-B, Table I"}],"minor_comments":[{"comment":"The phrase 'addictive cost functions' should be 'additive cost functions'.","section":"Section IV-B"},{"comment":"In line 4 of Algorithm 1, the displayed formula p_t(Y) = p_s(Y) * p_t(Y)/p_s(Y) is tautological; please state the BBSE estimation step explicitly as estimating the ratio and then renormalizing.","section":"Section IV-C, Algorithm 1"},{"comment":"The claim that theta close to 0.5 creates an 'overly compact cluster structure' is informal; no quantitative measure of cluster compactness or its effect on the target risk is provided.","section":"Section III-B and Section V-C"},{"comment":"The caption and surrounding text should clarify that the synthetic validation compares the sampling-domain risk with the source-domain risk, and that the source-domain risk is intended to represent epsilon_c' in Proposition 1.","section":"Figure 6"}],"recommendation":"reject","confidential_remarks":"The paper is within the journal's scope, but the central theoretical comparison is unsound and the empirical results undercut the headline claim. The missing supplementary proofs compound the problem. I recommend rejection; a resubmission would need a corrected comparison result between sampling and reweighting, and a re-evaluation against MOT on the reported benchmarks."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is a mixed bag: the IS2C construction—sampling new labeled data by mixing same-class source pairs according to estimated target proportions—is a fresh idea in PDA, and the O(n²) reformulation of Tensor Sinkhorn for the binary domain variable is a genuine, checkable speedup. If you work on OT-based alignment, the fast ETIC part is worth a look.\n\nThe soft spots are in the theory and the claimed empirical status. Proposition 1 is not the advertised guarantee. The paper assumes every component of the probabilistic output is convex; but if K components are nonnegative, convex, and sum to 1, each of them must be affine. The premise therefore yields equality, not a smaller risk. The stress-test's counterexample is off—it computes a deterministic threshold error instead of the paper's randomized 0-1 loss, and for affine classifiers the mixture mean is unchanged, so the risks actually match—but the core point stands: the convexity assumption collapses to the linear case, and the paper's 'strictly smaller' claim for non-linear networks rests on experiments, not on Proposition 1. Theorem 1's proof is in the missing supplementary file, so I couldn't check it; as stated it is a plausible decomposition, but the label-shift term is zero by construction since Pc(Y)=Pt(Y) exactly.\n\nThe empirical section overstates things. In Table I, MOT beats IS2C on Office-Home (80.6 vs 79.2), VisDA (92.4 vs 89.3), and Image-CLEF (93.6 vs 93.2); IS2C only edges it on Office-31. The text says on VisDA IS2C is 'only lower than SLM', which is false—MOT is higher. Claims of 'superior performance over existing methods' need to be toned down to 'competitive'.\n\nThat said, the ablation on sampling vs reweighting (IS vs wERM with the same ETIC alignment) is the right comparison and shows real gains, and the ETIC complexity reduction is solid arithmetic. The paper is salvageable with major revision: fix Proposition 1 to say what it actually proves, provide the supplementary, and re-frame the empirical claims.\n\nRecommendation: send to peer review—it deserves a serious referee—but expect a reject-or-major-revision outcome.\n\nBest.","headline":"A promising sampling-based PDA method with a real O(n²) ETIC speedup, but Proposition 1 proves far less than claimed and the empirical SOTA claims are contradicted by the paper's own tables.","tokens_in":23557,"tokens_out":7792,"would_cite":false,"duration_ms":73358,"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":"Sampling new labeled data from a constructed distribution, rather than reweighting old source samples, controls target risk in partial domain adaptation through a four-term error bound.","keywords":["partial domain adaptation","importance sampling","generalization error analysis","label shift","conditional shift","optimal transport","independence criterion","importance weighting"],"falsifier":"Train a fixed non-convex network on a fixed source domain and compare the empirical risk on the sampled domain $P_c$ with the empirical risk on the weighted-source domain for the same target label proportions; if the sampling-domain risk ever exceeds the weighted-source risk after convergence, the claimed dominance of sampling over reweighting beyond the convex case is falsified. The paper's own Figures 6(c) and 6(d) cover only two real-data checkpoints, so a systematic sweep over tasks and seeds can settle it.","tokens_in":22396,"feed_emoji":"🎯","tokens_out":8627,"duration_ms":77743,"temperature":0.7,"pith_summary":"This paper makes the case that label-shift correction in partial domain adaptation does not have to happen by reweighting source samples. Instead, it proposes to build a new 'sampling domain': draw pairs of same-class source samples, mix them with ratio $\\theta$, and use the mixture as fresh labeled data with the target label proportions. The central theoretical claim is that training on this sampling domain controls target risk, because the gap between sampling and target errors is bounded by a label-shift term, a conditional-shift term, and a mixing penalty. A second proposition says that, when model outputs are convex, the sampling-domain model has no larger risk than the source-domain or reweighted model. If these claims hold, importance sampling becomes a principled replacement for importance weighting, with the practical benefit that outlier classes vanish from training and the model can exploit synthetic but labeled data.","feed_headline":"Why sampling beats reweighting for partial domain adaptation","feed_subtitle":"Mixing same-class source samples to match target label proportions gives a provable target-risk bound and better accuracy.","key_machinery":"The machinery is the sampling distribution $P_c$: for each class $j$, draw two source samples $X_1,X_2$ from $P_s(X\\mid Y=j)$ and form $X=\\theta X_1+(1-\\theta)X_2$, while setting $p_c(Y=j)=p_t(Y=j)$. Theorem 1 is the identity that carries the argument, decomposing the sampling-to-target error gap into the balanced prediction error, the conditional error gap $\\Delta_{CE}$, and the mixing penalty $2\\ell(K-1)\\sqrt{\\theta(1-\\theta)}\\,C_s$. Proposition 1 then compares $P_c$ with the degenerate cases $\\theta=0,1$ under convexity, showing the sampling-domain risk is no larger. The alignment side of the method is ETIC, an entropy-regularized optimal transport independence criterion, whose empirical estimator is accelerated from $\\mathcal{O}(n_j^3)$ to $\\mathcal{O}(n_j^2)$ per iteration by using the discreteness of the domain variable.","core_discovery":"On its own terms, the paper claims that a model trained on the sampled domain $P_c$ --- built from same-class convex mixtures $X = \\theta X_1 + (1 - \\theta) X_2$ with $p_c(Y)=p_t(Y)$ --- controls the target risk through the inequality $$|\\varepsilon_t(h\\circ g) - \\varepsilon_c(h\\circ g)| \\le \\|P_c(Y)-P_t(Y)\\|_1 \\Delta_{BE}(P_c)(\\hat Y \\| Y) + (K-1)\\Delta_{CE}(\\hat Y) + 2\\ell(K-1)\\sqrt{\\$\\theta$(1-\\$\\theta$)}C_s.$$ Each term is interpretable: the risk on the sampling domain, the label-distribution gap times the balanced prediction error, the class-conditional gap across shared classes, and a penalty from mixing. Under the additional condition that every component of $h \\circ g$ is convex in the input, Proposition 1 gives $\\varepsilon_c(h\\circ g) \\le \\varepsilon_{c'}(h\\circ g)$, where $P_{c'}$ is the unmixed source sampling domain, so the mixture never hurts and typically helps. The method then estimates the target label proportions with BBSE, samples from $P_c$, and trains by minimizing the classification loss on the sampled data together with an optimal-transport independence criterion (ETIC) for class-conditional alignment. Experiments across Office-31, Office-Home, Image-CLEF and VisDA-2017 are reported as consistent with these bounds, with accuracy gains over reweighting baselines.","pith_inferences":["Editorial inference: the same mixture construction is a natural template for open-set domain adaptation, where the target includes unknown classes; one would need to allocate a mass to an 'unknown' class rather than deleting outliers, a direction the paper names as future work.","Editorial inference: because the bound's mixing penalty scales with $\\sqrt{\\theta(1-\\theta)}$ and the source radius $C_s$, normalizing or whitening features before sampling could tighten the bound in high-dimensional feature spaces; the paper does not test this.","Editorial inference: if Proposition 1's convexity is genuinely unnecessary, then a direct re-derivation through interpolation or Jensen-type inequalities may yield a weaker but more general dominance statement; the paper's evidence is only two real-data checkpoints in Figures 6(c) and 6(d)."],"forward_implications":["If Theorem 1 holds, then PDA training can be reframed as risk minimization on a synthetic but labeled domain, so outlier classes drop out of the training set rather than lingering with small weights.","The four-term bound makes the mix-ratio $\\theta$ a directly tunable trade-off: $\\theta$ near 0 or 1 removes the mixing penalty but leaves cluster structure loose, while $\\theta$ near 0.5 tightens clusters but adds penalty, predicting the observed U-shaped sensitivity.","With the fast ETIC computation, class-conditional alignment costs $\\mathcal{O}(n^2)$ per iteration instead of $\\mathcal{O}(n^3)$, making the method practical for larger target batches in PDA.","The BBSE-based estimate of $p_t(Y)$ means the method does not require known target label proportions; it only needs pseudo-labels from the current model.","If Proposition 1 extends beyond convex models, sampling-based correction should dominate importance weighting in accuracy across many PDA benchmarks, not just the four reported datasets."],"supporting_citations":[{"why":"Defines the balanced prediction error and conditional error gap that Theorem 1 extends from shared-support label shift to partial domain adaptation.","marker":"[22]"},{"why":"Supplies the entropy-regularized optimal transport independence criterion and the Tensor Sinkhorn estimation used for class-conditional alignment.","marker":"[49]"},{"why":"Supplies the black-box shift estimation algorithm used to estimate the target label proportions $p_t(Y)$.","marker":"[54]"},{"why":"Supports the claim that mixing same-class samples reduces class-conditional variance, motivating the choice $\\theta \\in (0,1)$.","marker":"[48]"},{"why":"Supplies the symmetric Sinkhorn divergence that underlies the ETIC independence measure.","marker":"[52]"},{"why":"Provides the reweighting-based partial adversarial domain adaptation baseline whose weighting strategy IS2C replaces.","marker":"[28]"}],"fun_headline_variants":["Sampling beats reweighting for partial domain adaptation","Mix source samples to fix label shift in partial DA","IS²C: provable shift correction without reweighting","For partial DA, sample don't reweight"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the model's output is convex in the input; the paper itself calls this assumption 'quite strong,' and the neural network used in experiments is not convex.","fun_headline_variants_meta":{"raw":{"variants":["Sampling beats reweighting for partial domain adaptation","Mix source samples to fix label shift in partial DA","IS²C: provable shift correction without reweighting","For partial DA, sample don't reweight"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0003,"raw_usage":{"total_tokens":1837,"prompt_tokens":1157,"completion_tokens":680,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":773,"completion_tokens_details":{"reasoning_tokens":616}},"tokens_in":773,"tokens_out":680,"duration_ms":6684,"temperature":1.0,"reasoning_tokens":616,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:48:49.368325+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a fixed non-convex network on a fixed source domain and compare the empirical risk on the sampled domain $P_c$ with the empirical risk on the weighted-source domain for the same target label proportions; if the sampling-domain risk ever exceeds the weighted-source risk after convergence, the claimed dominance of sampling over reweighting beyond the convex case is falsified. The paper's own Figures 6(c) and 6(d) cover only two real-data checkpoints, so a systematic sweep over tasks and seeds can settle it.","supporting_citations":[{"cited_title":"Domain adaptation with conditional distribution matching and generalized label shift,","cited_arxiv_id":null,"evidence_quote":"Defines the balanced prediction error and conditional error gap that Theorem 1 extends from shared-support label shift to partial domain adaptation."},{"cited_title":"Entropy regularized optimal trans- port independence criterion,","cited_arxiv_id":null,"evidence_quote":"Supplies the entropy-regularized optimal transport independence criterion and the Tensor Sinkhorn estimation used for class-conditional alignment."},{"cited_title":"Detecting and correcting for label shift with black box predictors,","cited_arxiv_id":null,"evidence_quote":"Supplies the black-box shift estimation algorithm used to estimate the target label proportions $p_t(Y)$."},{"cited_title":"On Mixup Regularization","cited_arxiv_id":"2006.06049","evidence_quote":"Supports the claim that mixing same-class samples reduces class-conditional variance, motivating the choice $\\theta \\in (0,1)$."},{"cited_title":"Interpolating between optimal transport and mmd using sinkhorn divergences,","cited_arxiv_id":null,"evidence_quote":"Supplies the symmetric Sinkhorn divergence that underlies the ETIC independence measure."},{"cited_title":"Partial adversarial domain adaptation,","cited_arxiv_id":null,"evidence_quote":"Provides the reweighting-based partial adversarial domain adaptation baseline whose weighting strategy IS2C replaces."}],"review_version":2}