{"id":"08f71c33-497a-44da-b5fe-b8f3f88f7a28","arxiv_id":"2502.08873","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"p-conductance learning solves an affine-relaxed p-norm mincut over diffused label measures, linking mincut, effective resistance, and Wasserstein distance, and reporting strong accuracy in low-label, corrupted-label, and partial-label graph classification.","lead":"Graph-based semi-supervised learning usually fails when very few or corrupted labels are available. The authors propose p-conductance learning, a family of minimum-cut problems on label probability measures, and report strong accuracy on citation and image benchmarks at low label rates.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The experimental definition of ϵ is internally contradictory: Section 4 says ϵ=0 means no cardinality prior and ϵ=n means exact class sizes, while Remark D.1 and Eq. (28) imply the reverse. The tables therefore do not identify which variant achieves the reported SOTA accuracy.","rationale":"The reader's verdict is CONDITIONAL, and the reader already noted the ϵ inconsistency in the rationale, though not as the weakest assumption. I agree with the reader's overall assessment: the gauge-duality core and the p=2 reduction to effective resistance are mathematically clean, and the proposed SSNAL algorithm is a plausible contribution. However, the experimental reporting has a concrete internal contradiction that is central to the paper's empirical claims. The main text and Appendix D.1 assign opposite meanings to ϵ=0 and ϵ=n, so the tables do not currently establish which method variant achieves the reported accuracy. This is not a matter of interpretation or a missing proof; it is a direct inconsistency between two parts of the manuscript. The Theorem 2.7 gap identified by the reader (potential perturbation vs. classification accuracy) is also real and limits the theoretical support for robustness, but it is a limitation rather than a contradiction in the reported experiments. Because the issue is fixable by relabeling or re-running, I do not recommend moving from CONDITIONAL to REJECT; the appropriate action is to require the authors to disambiguate ϵ and verify that the table labels match the implementation. The absence of code makes this verification impossible from the paper alone, reinforcing the need for a conditional verdict rather than full acceptance.","tokens_in":24259,"tokens_out":10276,"duration_ms":107536,"concrete_test":"Re-run the p-conductance experiments using Eq. (28) with ϵ=0 (exact class-size constraint) and ϵ=n (argmax thresholding), and compare the resulting accuracies to the rows labeled 'p-conductance (p=2, ϵ=0)' and 'p-conductance (p=2, ϵ=n)' in Tables 1–4. If the row currently labeled ϵ=0 is reproduced by the ϵ=n implementation, or if the accuracies swap when the correct definitions are used, the paper's reported SOTA comparison is mislabeled and the empirical claim requires revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline empirical claim depends on comparing the variants labeled 'p-conductance (ϵ=0)' and 'p-conductance (ϵ=n)', but the meaning of ϵ is stated inconsistently. Section 4 says: 'we denote by p-conductance (ϵ = 0) to be our method without any cardinality prior and p-conductance (ϵ = n) to be our method with exact knowledge of class cardinalities.' Yet Appendix D.1 defines the relaxation max ⟨ϕ,P⟩ subject to P ≥ 0, P1 = 1, m−ϵ ≤ P^T 1 ≤ m+ϵ, and states: 'When ϵ = 0, this is a linear program over the transportation polytope. It is known that the simplex method recovers integer solutions... When ϵ = n, the optimum corresponds to the standard heuristic of thresholding the largest corresponding component of the vector ϕ∗.' These two statements are exact opposites: ϵ=0 enforces exact cardinalities, and ϵ=n is the no-prior argmax rule. The numerical tables (e.g., Cora at 1 label per class: 63.1 vs 58.9) are therefore labeled inconsistently, so a reader cannot tell which configuration produced the reported numbers. Since no code or data is provided, this internal contradiction is not cosmetic: it directly undermines the abstract's claim of state-of-the-art accuracy in the low-label-rate regime. The l2-to-accuracy gap in Theorem 2.7 is a further limitation, but the ϵ inconsistency is more immediately load-bearing because it concerns the very experiments that establish the main empirical result.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes p-conductance learning, a graph-based SSL method that solves a measure-constrained p-Dirichlet energy minimization problem with an affine label constraint. It derives connections to mincut, randomized cuts, normalized cuts, effective resistance, and Wasserstein distance via gauge duality, proves an ℓ2 stability bound for diffused labels at p=2, develops a semismooth Newton augmented-Lagrangian algorithm, and reports experiments on citation and image datasets claiming state-of-the-art accuracy in low-label-rate, corrupted-label, and partial-label regimes.","tokens_in":24663,"tokens_out":8854,"duration_ms":91366,"significance":"If the claims hold, the paper gives a clean variational framework that unifies p-Laplace and Poisson learning with classical graph-cut and optimal-transport objects; the explicit p=2 minimizer in Eq. (8) and the gauge-duality relationships in Theorem 2.4 are attractive and appear correct. The SSNAL algorithm is a standard and well-cited extension of existing techniques. However, the empirical headline is currently not verifiable: the meaning of the cardinality parameter ε is stated inconsistently between Section 4 and Appendix D.1, the experimental configuration (p, t, ε) is under-specified, and no code or data is provided. In addition, Theorem 2.7 bounds the ℓ2 perturbation of the potential but does not imply sign or argmax preservation, so the theoretical support for the robustness claim is weaker than the abstract suggests.","major_comments":[{"comment":"Section 4 states: 'we denote by p-conductance (ε = 0) to be our method without any cardinality prior and p-conductance (ε = n) to be our method with exact knowledge of class cardinalities.' Appendix D.1, Eq. (28), relaxes the cardinality constraint to m−ε ≤ Pᵀ1 ≤ m+ε, and Remark D.1 states that ε=0 is the LP over the transportation polytope with exact cardinalities, while ε=n corresponds to the unconstrained argmax heuristic. These two statements are exact opposites. Because the tables (e.g., Table 1, Cora with 1 label per class, 63.1 for ε=0 versus 58.9 for ε=n) use these labels without code or data, a reader cannot determine which configuration produced the reported state-of-the-art numbers. The convention must be fixed and the experimental tables re-labeled or re-run accordingly.","section":"Section 4 and Appendix D.1"},{"comment":"Theorem 2.7 bounds ∥ψ − ψ~t∥₂ in terms of the perturbation η and the diffusion time t, but the method's predictions are sgn(φ*) or argmax over columns of φ*. An ℓ2 bound on the potential does not imply that signs or argmax orderings are preserved; no margin or separation assumption is stated. Moreover, Remark 2.8 shows that the improvement condition requires ∥η∥₂/∥µ−ν∥₂ > 1, which in the binary example means more than one quarter of the labels are corrupted, so the theorem does not cover the small-corruption regime shown in the experiments. The diffusion time t is also chosen empirically in Figure 3 rather than prescribed by the theorem. As a result, the theoretical result only motivates, rather than establishes, the abstract's robustness claim, and the gap between the ℓ2 statement and classification accuracy should be addressed explicitly.","section":"Section 2.2, Theorem 2.7"},{"comment":"The experimental configuration is under-specified. The values of p, the diffusion time t, and the cardinality slack ε are not reported for each dataset and label rate; only the graph construction parameter k=10 is fixed for image datasets. Furthermore, for p=2, Eq. (8) shows that the optimal potential is a positive scalar multiple of the Poisson learning potential L†(µ−ν). Consequently, the gains reported for p=2 over the Poisson baseline could come entirely from the heat-kernel diffusion and the cardinality-aware rounding in Eq. (28), rather than from the new variational objective. An ablation separating these components (for example, Poisson learning with diffused labels, and p=2 with and without cardinality rounding) is needed to support the attribution in the abstract, and the absence of code or data makes these variants impossible to reconstruct.","section":"Section 4, Tables 1–6"}],"minor_comments":[{"comment":"The convergence-rate statement is missing a closing parenthesis; it should read ‖φ^{k+1} − φ̄‖ = O(‖φ^k − φ̄‖^{1+τ}).","section":"Theorem 3.7"},{"comment":"The proposition states 'integer values of p ∈ [0, ∞)', but the problem (Cp) is defined for p ≥ 1; p=0 is not used anywhere.","section":"Proposition 3.5"},{"comment":"The CIFAR-100 PLL table lists 'p-conductance (p = 2, ε = n)' twice with different accuracy values; one of these rows should presumably be labeled with the other ε value used in the paper.","section":"Table 6"},{"comment":"Equation (29) writes 'φ∗_i = arg max ...', but the left-hand side should be the predicted class label, not the optimal potential.","section":"Appendix D.1, Eq. (29)"},{"comment":"The text says 'p-conductance learning outperforms all other methods' in the image-dataset section, but on MNIST p-conductance-MBO ties PoissonMBO exactly at every reported label count (97.5, 97.5, 97.5, 97.6, 98.1); the claim should be softened to 'matches or outperforms'.","section":"Section A.1, Table 2"}],"recommendation":"major_revision","confidential_remarks":"The ε inconsistency is the most serious issue because it directly affects the interpretation of every table and the abstract's SOTA claim. The p=2 reduction to a rescaled Poisson potential should also be acknowledged more explicitly, since it shifts the claimed novelty to the diffusion and rounding steps. If the authors correct the notation, supply the missing experimental configuration, and clarify the relationship between Theorem 2.7 and classification accuracy, the paper would be a solid contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: the theoretical core is genuinely worth your time, but the experimental section has a self-contradictory definition of epsilon that makes the headline SOTA claims uninterpretable as written. I'd send it to review anyway, with the empirical claims on probation.\n\nWhat's new: the affine-relaxed measure mincut family C_p (and the p=∞ variant) is a clean object that connects p-Laplacian regularization, mincut, effective resistance, and Wasserstein distance through gauge duality. The p=2 minimizer is exactly a rescaled Poisson learning potential, so the predictor direction itself isn't new; the family and the duality relations are. The SSNAL/SSNCG algorithm is a competent application of known augmented Lagrangian and semismooth Newton machinery; the convergence theorems follow from Rockafellar and Zhao et al., and the proofs look standard and correct. Theorem 2.7 is also correct as a bound on the potential, and the observation that diffusion helps when label noise is large relative to the clean signal is neat.\n\nThe soft spots are load-bearing. Section 4 defines ϵ=0 as 'no cardinality prior' and ϵ=n as 'exact class sizes.' Remark D.1 and Eq. (28) define the opposite: ϵ=0 is the transportation polytope with exact sizes, and ϵ=n collapses to the argmax threshold. The tables report both variants (Cora 63.1 vs 58.9), so the reader cannot tell which configuration produced the reported numbers. No code or data is provided to disambiguate. This is not a cosmetic notation clash; it directly undermines the abstract's robustness and SOTA claims. Also, Theorem 2.7 bounds the ℓ2 perturbation of the potential L⁺(μ−ν); it does not imply sign or argmax stability, and the diffusion time t is chosen empirically in Figure 3. So the theoretical robustness guarantee does not directly transfer to classification accuracy.\n\nWho this is for: researchers in graph SSL theory, p-Laplacian methods, and gauge duality will get real value from the first half. The empirical section needs to be re-run with consistent definitions and released code/data before anyone should build on those numbers. Recommendation: send to peer review, because the theoretical contribution deserves serious referees, but make acceptance conditional on fixing the epsilon definitions, supplying code/data, and re-reporting the tables with unambiguous labels.","headline":"The gauge-duality core and p-conductance family are a real contribution, but the experimental section's self-contradictory epsilon definitions make the headline accuracy claims unreadable as written.","tokens_in":25150,"tokens_out":4348,"would_cite":false,"duration_ms":37332,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68Q87","90C35","05C90","90C53"],"pacs":[],"model":"deepseek-v4-flash","headline":"A family of p-conductance energy programs, solved with heat-kernel diffused labels, is claimed to achieve state-of-the-art accuracy in graph-based semi-supervised learning under scarce, corrupted, and partial labels.","keywords":["semi-supervised learning","node classification","p-Laplacian","graph effective resistance","optimal transport","label corruption","heat kernel","measure mincut"],"falsifier":"Take a graph and label noise $\\eta$ satisfying the Theorem 2.7 condition $\\|\\eta\\|_2>\\|\\mu-\\nu\\|_2$, and compute the fraction of nodes where sign($\\phi$) changes between the clean and diffused corrupted potentials. If a small $\\ell^2$ perturbation flips the sign at a large fraction of high-confidence nodes, or if increasing $t$ decreases accuracy rather than increasing it, the robustness mechanism described in the paper is not what drives the empirical gains.","tokens_in":24080,"feed_emoji":"⚡","tokens_out":5637,"duration_ms":45629,"temperature":0.7,"pith_summary":"This paper proposes p-conductance learning, a graph-based semi-supervised method that replaces hard label constraints with an affine constraint on two label probability measures. The central object is the energy program $C_p(\\mu,\\nu)=\\min_{\\phi:\\phi^T(\\mu-\\nu)=1}(\\sum_{ij} w_{ij}|\\phi_i-\\phi_j|^p)^{1/p}$, which interpolates between sparse-cut behavior at $p=1$, effective-resistance behavior at $p=2$, and Wasserstein/transport behavior at $p=\\infty$. The authors argue that diffusing the labels with the heat kernel before solving makes the resulting sign predictions robust to label corruption, and they support this with a stability bound for $p=2$ plus experiments on citation and image datasets. Sympathetically read, the paper's claim is that one tunable energy family, solved by a fast semismooth Newton method, achieves state-of-the-art accuracy in exactly the regimes where labels are scarce, noisy, or only partially known.","feed_headline":"p-conductance learning tops low-label and noisy-label tests","feed_subtitle":"A single p-norm energy with heat-diffused labels generalizes Poisson learning and wins on citation and image benchmarks.","key_machinery":"The load-bearing object is the measure p-conductance $C_p(\\mu,\\nu)$, a variational program in which the unknown potential $\\phi$ is coupled to the two label measures through the affine constraint $\\phi^T(\\mu-\\nu)=1$ instead of hard node-wise label equalities. The paper shows via gauge duality that $C_p$ is the reciprocal of a p-Beckmann optimal-transport metric, which yields the effective-resistance identity at $p=2$ and the Wasserstein identity at $p=\\infty$. Robustness is carried by the heat-kernel diffusion operator $e^{-tL}$ applied to the label measures, combined with a spectral bound on $L^+$; Theorem 2.7 bounds the $\\ell^2$ change in the potential by $t\\|\\mu-\\nu\\|_2 + \\lambda^{-1} e^{-t\\lambda} \\|\\eta\\|_2$. Computationally, the paper introduces a semismooth Newton conjugate-gradient augmented-Lagrangian method for all $p\\in[1,\\infty]$, whose convergence rests on strong semismoothness of the proximal map of the weighted p-norm.","core_discovery":"On the paper's own terms, the discovery is that the measure p-conductance program $C_p$ — an affine-relaxed, p-Laplacian-style objective — is the right interpolation between Poisson learning and p-Laplace learning, and that heat-kernel diffused labels make it robust. For $p=2$ the minimizer is a rescaling of the Poisson potential, $\\phi^*=L^+(\\mu-\\nu)/((\\mu-\\nu)^T L^+ (\\mu-\\nu))$, so the method literally generalizes Poisson learning. The paper proves that $C_1$ is a generalized min-cut/max-flow problem, that $C_2^2$ is the reciprocal of measure effective resistance, and that $C_\\infty$ is the reciprocal of a 1-Wasserstein distance. It also proves that for $p=2$, applying $e^{-tL}$ to the labels improves a worst-case $\\ell^2$ bound on the potential when label noise is large relative to the clean signal. Empirically, the method reports higher accuracy than Poisson learning and PoissonMBO in low label-rate, corrupted-label, and partial-label settings on Cora, Citeseer, Pubmed, MNIST, FashionMNIST, CIFAR-10, and CIFAR-100.","pith_inferences":["The stability bound in Theorem 2.7 concerns the $\\ell^2$ distance of potentials, not the sign or argmax of predictions; a natural extension is to prove that, under a margin or separation condition on the clean potential, the same diffusion step also stabilizes the label error rate, which would close the gap between the theorem and the empirical robustness claims.","Because $C_\\infty$ is the reciprocal of a shortest-path Wasserstein distance, the p-conductance family can be read as a one-parameter interpolation between cut geometry and transport geometry; this suggests testable connections between label robustness and the coarse geometry of the graph, e.g., graphs with larger spectral gap should show smaller accuracy drops under noise.","The diffusion time $t$ is treated as a free parameter and chosen empirically in Figure 3; one could derive a data-dependent rule for $t$ from the bound in Theorem 2.7, e.g., $t\\approx(1/\\lambda)(\\|\\eta\\|/\\|\\mu-\\nu\\| - 1)$, and test whether the automatically chosen $t$ matches the empirical optimum."],"forward_implications":["At $p=2$ with no cardinality prior ($\\epsilon=0$), p-conductance learning reports the best or tied-best accuracy on Cora and Pubmed across label rates from 1 to 100 labels per class, including a 4.6% gain over PoissonMBO at 1 label per class on Cora.","On CIFAR-10 with 40% flipped labels, the method reports a 3.8% accuracy improvement over Poisson learning, and the experiments show the gain depends on heat-kernel diffusion time.","In the CIFAR-100 partially-labeled setting with a candidate set of size 4, p-conductance-MBO outperforms PoissonMBO by 4.1%.","The same energy family covers $p=1,2,\\infty$, giving a unified view: sparse cuts at $p=1$, effective resistance at $p=2$, and transport geometry at $p=\\infty$."],"supporting_citations":[{"why":"Defines Laplace learning, the harmonic-extension baseline that this paper generalizes and against which it benchmarks low label-rate performance.","marker":"Zhu et al. (2003)"},{"why":"Introduces Poisson learning, the method whose objective and MBO variant p-conductance learning generalizes and outperforms in the experiments.","marker":"Calder et al. (2020)"},{"why":"Develops p-Laplace learning for general p, the family that motivates the p-energy in (Cp) and supplies the comparison results at low label rates.","marker":"Flores et al. (2022)"},{"why":"Provides the gauge duality theory used to prove the reciprocal relationship between Cp and the p-Beckmann/optimal-transport metrics.","marker":"Friedlander et al. (2014)"},{"why":"Supplies the augmented-Lagrangian convergence theory on which Theorem 3.6 rests.","marker":"Rockafellar (1976a,b)"},{"why":"Provides the semismooth Newton-CG convergence results used to establish Theorem 3.7.","marker":"Zhao et al. (2010)"},{"why":"Connects the Beckmann metric to the 1-Wasserstein distance, used in Corollary 2.5.","marker":"Peyré et al. (2019)"}],"fun_headline_variants":["p-conductance learning beats low-label and noisy-label baselines","Heat-diffused p-conductance generalizes Poisson learning","p-conductance unifies min-cut, resistance, and Wasserstein","Robust SSL via p-conductance: wins on corrupted labels"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes that a small $\\ell^2$ perturbation of the continuous potential (bounded in Theorem 2.7) translates into unchanged sign or argmax predictions; the link between that bound and classification accuracy is not proved, and the diffusion time $t$ is chosen empirically.","fun_headline_variants_meta":{"raw":{"variants":["p-conductance learning beats low-label and noisy-label baselines","Heat-diffused p-conductance generalizes Poisson learning","p-conductance unifies min-cut, resistance, and Wasserstein","Robust SSL via p-conductance: wins on corrupted labels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000514,"raw_usage":{"total_tokens":2513,"prompt_tokens":981,"completion_tokens":1532,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":597,"completion_tokens_details":{"reasoning_tokens":1457}},"tokens_in":597,"tokens_out":1532,"duration_ms":11911,"temperature":1.0,"reasoning_tokens":1457,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T23:22:45.912794+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a graph and label noise $\\eta$ satisfying the Theorem 2.7 condition $\\|\\eta\\|_2>\\|\\mu-\\nu\\|_2$, and compute the fraction of nodes where sign($\\phi$) changes between the clean and diffused corrupted potentials. If a small $\\ell^2$ perturbation flips the sign at a large fraction of high-confidence nodes, or if increasing $t$ decreases accuracy rather than increasing it, the robustness mechanism described in the paper is not what drives the empirical gains.","supporting_citations":[{"cited_title":"Semi-supervised learning using Gaussian fields and harmonic functions","cited_arxiv_id":null,"evidence_quote":"Defines Laplace learning, the harmonic-extension baseline that this paper generalizes and against which it benchmarks low label-rate performance."},{"cited_title":"Poisson learning: Graph based semi-supervised learning at very low label rates","cited_arxiv_id":null,"evidence_quote":"Introduces Poisson learning, the method whose objective and MBO variant p-conductance learning generalizes and outperforms in the experiments."},{"cited_title":"Analysis and algorithms for p-based semi-supervised learning on graphs","cited_arxiv_id":null,"evidence_quote":"Develops p-Laplace learning for general p, the family that motivates the p-energy in (Cp) and supplies the comparison results at low label rates."},{"cited_title":"P., Macedo, I., and Pong, T","cited_arxiv_id":null,"evidence_quote":"Provides the gauge duality theory used to prove the reciprocal relationship between Cp and the p-Beckmann/optimal-transport metrics."},{"cited_title":"A newton-cg augmented lagrangian method for semidefinite programming","cited_arxiv_id":null,"evidence_quote":"Provides the semismooth Newton-CG convergence results used to establish Theorem 3.7."}],"review_version":1}