{"id":"f71d812b-af69-4590-8b98-6b8baee4508f","arxiv_id":"2506.04669","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"PML-FSMIR reconstructs candidate labels with a mutual-information matrix and reweights selected features by label connectivity, scoring first on most benchmark metrics in the reported experiments.","lead":"This paper introduces PML-FSMIR, a three-stage feature selection method for partial multi-label learning, where each example carries several candidate labels and some are wrong. It cleans noisy labels using label relationships, then reweights features toward the most connected labels, and reports large F1 gains over eight baselines on benchmark datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claimed 'first in all but two cases' is contradicted by the paper's own Table 3: Coverage on LLOGF and CHD49 are not first, so the central superiority claim is unsupported.","rationale":"The paper's central claim is comparative superiority, so the experimental tables are the load-bearing evidence. The strongest statement of that evidence, Section 4.3, is contradicted by the paper's own Table 3: Coverage on LLOGF (PML-FSMIR 0.53 vs. PML-FP 0.48 and PML-LC 0.51) and Coverage on CHD49 (PML-FSMIR 0.50 vs. PML-FSSO and DRMFS 0.49) place PML-FSMIR third, not second. This directly falsifies the headline claim that only two exceptions exist and that both are second-place finishes. The reader's verdict of REJECT remains appropriate, and I would not change it. My primary concern differs from the reader's stated weakest assumption (noise-resistance of the mutual-information matrix); I focus on the internally inconsistent ranking claim. The update-equation problem identified by the reader is also serious, but the ranking discrepancy is the most direct threat to the central claim as written. A typographical correction or code release could not repair the contradiction without changing the reported numbers.","tokens_in":14121,"tokens_out":7858,"duration_ms":81073,"concrete_test":"Recompute the per-dataset-per-metric rank of PML-FSMIR from the mean values in Tables 2-6 (Ranking Loss and Coverage: smaller is better; Average Precision, Macro-F1, Micro-F1: larger is better), including ties. Count the cells where PML-FSMIR is not strictly first. If the count exceeds 2, or if any asserted 'second' placement is actually third or worse, the Section 4.3 claim is false. This check requires no new experiments, only arithmetic on the published tables.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Sec. 4.3) states that PML-FSMIR 'ranks first in all cases except Coverage on CHD49 and Average Precision on LLOGF, while in these two cases PML-FSMIR all ranks second.' This is not what Tables 2-6 report. In Table 3 (Coverage, lower is better), the LLOGF row gives PML-FSMIR 0.53, PML-LC 0.51, and PML-FP 0.48, so PML-FSMIR ranks third, not second. In the CHD49 row, PML-FSMIR is 0.50 while PML-FSSO and DRMFS are both 0.49, again third. The prose explicitly says 'all ranks second' in both of these cases, but the reported means rank third. Because the paper's quantitative evidence for 'superiority' is this ranking claim, and the claim is false as written, the central contribution is not supported by the reported experiments.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PML-FSMIR, a three-stage feature selection method for partial multi-label learning (PML). Stage 1 reconstructs the candidate label matrix using a mutual-information matrix to suppress label noise (Eq. 2). Stage 2 learns a feature weight matrix through a 'reformed low-rank' objective that includes a graph-Laplacian regularization term on the weight matrix (Eq. 7). Stage 3 reconstructs the weight matrix using label-connectivity information to emphasize features that are important for representative labels (Eq. 12). The method is evaluated on eight benchmark datasets against eight existing methods, and the paper reports ablation studies and parameter sensitivity analyses.","tokens_in":14376,"tokens_out":6468,"duration_ms":63965,"significance":"The general idea of using stable label relationships, via mutual information, for noise-resistant feature selection in PML is interesting and the authors provide a code repository and extensive experiments. If the method worked exactly as claimed, it would be a useful contribution to PML feature selection. However, the manuscript in its current form contains load-bearing inconsistencies: the optimization update for W does not match the stated objective, the claim of breaking free of the low-rank assumption is contradicted by the model itself, and the central empirical superiority claim is contradicted by the paper's own tables. These issues prevent a reliable assessment of the method's contribution.","major_comments":[{"comment":"The multiplicative update for W omits the graph Laplacian term that is central to the objective in Eq. (7). Differentiating β Tr(W L_T W^T) with respect to W yields a term proportional to L_T W (or W L_T), but the denominator in Eq. (11) contains β W instead. Furthermore, the parameter δ in Eq. (11) is never defined; presumably it should be related to γ from the l2,1-norm relaxation, but this is not stated. Consequently, the iterative procedure described in Algorithm 1 does not minimize the stated objective, and the reported experimental results cannot be attributed to the method as specified.","section":"3.2, Eq. (11)"},{"comment":"The paper repeatedly claims that the method 'breaks free of the low-rank assumption' and 'preserves the dimensions of the sample space.' This is contradicted by Eq. (4), which contains α||X - U V||^2_F with U ∈ R^{n×k} and V ∈ R^{k×d}. This is exactly a rank-k approximation of X, and the first term uses U V, which is a low-rank product. The formulation therefore still relies on a low-rank factorization of the feature matrix; the claim of avoiding the low-rank assumption is a conceptual overstatement and is not supported by the model.","section":"3.2, Eq. (4); Introduction and Contributions"},{"comment":"The key empirical claim that PML-FSMIR 'ranks first in all cases except Coverage on CHD49 and Average Precision on LLOGF, while in these two cases PML-FSMIR all ranks second' is false. In Table 3 (Coverage, lower is better), on LLOGF PML-FSMIR has 0.53 while PML-LC has 0.51 and PML-FP has 0.48, so PML-FSMIR ranks third, not second. On CHD49 in the same table, PML-FSMIR has 0.50 while PML-FSSO and DRMFS both have 0.49, again ranking third. The prose also omits Coverage on LLOGF as an additional case where the method is not first. Because this ranking statement is the paper's central evidence of superiority, the quantitative support for the main contribution is lacking.","section":"4.3, Tables 3 and 4"},{"comment":"The expression T_{ij} = y_{ij} Σ_{k=1}^q y_{ik} I(y_{ij}, y_{ik}) uses the notation I(y_{ij}, y_{ik}), which is undefined: mutual information is a function of two random variables (label columns), not of individual entries. The intended meaning is presumably the mutual-information matrix entry Z_{jk} = I(y_{:j}, y_{:k}). This notational error makes the derivation of the label reconstruction matrix difficult to follow and should be corrected.","section":"3.1, Eq. (3)"},{"comment":"The noise-resistance premise of the first stage is not adequately tested. The mutual information matrix Z in Eq. (1) is computed from the observed candidate label matrix Y, which already contains noise, and the reconstruction T in Eq. (2) assumes that label co-occurrence statistics are stable under noise. The experiments use only a fixed 20% noise level and provide no validation against clean ground-truth labels. Without varying the noise level or otherwise demonstrating that the estimated Z preserves the true label structure under corruption, the claimed 'noise-resistant' property of the label reconstruction stage is not established.","section":"3.1 and 4.5"}],"minor_comments":[{"comment":"The terms 'Marco-F1' and 'Mirco-F1' appear throughout; these should be 'Macro-F1' and 'Micro-F1'.","section":"4.3 and Tables 5-6"},{"comment":"The text states that Figure 5 shows parameter sensitivity on LLOGF in terms of Ranking Loss, but the figure caption says 'Parameter sensitivity studies on the CAL in terms of Coverage.' The x-axis labels in the figure are also garbled ('0.001 10008% 20%').","section":"4.4, Figure 5"},{"comment":"The text refers to 'Figure ??' when discussing the per-dataset curves; this should be a concrete figure number, presumably Figure 4.","section":"4.3"},{"comment":"The loop condition 'while not coverage do' should be 'while not converged do'.","section":"Algorithm 1, line 4"},{"comment":"The graph Laplacian regularization term is written inconsistently across equations: 'β Tr(W )LT (W )T' in Eq. (6), 'β Tr(W )LT (W )T' in Eq. (7), and 'β TrW L_T W^T' in Eq. (8). The intended form is Tr(W L_T W^T), and the notation should be uniform.","section":"3.2, Eqs. (6)-(8)"},{"comment":"The citation of Beck and Teboulle [2009] for the multiplicative gradient descent strategy is inaccurate; that reference introduces FISTA, not multiplicative update rules. A more appropriate reference would be to multiplicative non-negative matrix factorization updates.","section":"3.2, after Eq. (8)"}],"recommendation":"reject","confidential_remarks":"The manuscript has serious internal inconsistencies that go beyond presentation. The W update in Eq. (11) does not optimize the stated objective, the 'low-rank-free' claim is contradicted by the model's own factorization, and the headline ranking claim in Section 4.3 is false when checked against Table 3. These issues are load-bearing: the experimental results cannot be attributed to the method as described, and the empirical evidence for superiority is not credible. The paper would need substantially reworked derivations, corrected experiments, and revised claims before it could be reconsidered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version. PML-FSMIR is a genuinely new three-stage pipeline for partial multi-label feature selection, and the reported F1 gains over eight baselines are big—on Macro-F1 and Micro-F1 the margins are often 2–4x the next best method. But the paper as written has three concrete problems that a referee would have to flag before the empirical results can be trusted.\n\nThe first is formal. The W update in Eq. 11 is not the minimizer of the stated objective in Eq. 7: it drops the graph Laplacian L_T that appears in the manifold term, and it introduces an undefined δ. The derivative of βTr(W L_T W^T) is 2βW L_T, but Eq. 11 has βW, no L_T. That means the iterative procedure does not optimize the model the paper describes. This is fixable, but it needs to be fixed or justified.\n\nSecond, the \"breaks free of the low-rank assumption\" claim is overstated. The objective still includes α||X - UV||²_F, which is exactly a low-rank factorization of X; U V has rank ≤ k. The method reformulates the low-rank assumption, it does not abandon it. The abstract and contributions should say that.\n\nThird, the headline ranking claim is misstated. The paper says PML-FSMIR ranks first everywhere except Coverage on CHD49 and Average Precision on LLOGF, and \"ranks second\" in those two cases. The tables support the first part, but in Coverage on CHD49 the method's 0.50 is behind PML-FSSO and DRMFS (both 0.49), and on LLOGF its 0.53 is behind PML-LC (0.51) and PML-FP (0.48). In both cases it ranks third, not second. The overall superiority claim still holds, but the text needs a correction.\n\nThere are also two softer issues. The first-stage premise—that mutual information computed from noisy candidate labels is stable enough to reconstruct clean labels—is tested at a single 20% noise level and only by ablation; there is no sweep over noise levels or check against clean labels. And the parameter sensitivity figure caption says CAL/Coverage while the text says LLOGF/Ranking Loss, a minor mismatch.\n\nWhat I like: the method is new relative to the cited literature, the code is released, the ablation clearly shows both added stages matter, and the F1 improvements are large enough that if the optimization is correct, this is a meaningful advance for PML feature selection. The citation pattern looks fine to me; I do not see self-citation inflation.\n\nBottom line: this deserves a serious referee. The idea and empirical results are compelling enough to warrant a major-revision cycle, not a desk reject. But the update-equation error and the misstated ranking have to be addressed, and the noise-resistance claim needs stronger evidence.","headline":"PML-FSMIR is a genuinely new three-stage pipeline with large reported F1 gains, but the paper's update equations don't match its stated objective and its headline ranking claim is misstated — both fixable, but not ignorable.","tokens_in":14862,"tokens_out":5205,"would_cite":false,"duration_ms":56217,"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":"This paper claims that a three-stage method for partial multi-label learning, PML-FSMIR, cleans the noisy candidate-label matrix using label-to-label mutual information, learns feature weights without reducing the data dimension, and then…","keywords":["partial multi-label learning","feature selection","mutual information","label noise","label reconstruction","label connectivity","positive label identification","sparse learning"],"falsifier":"Construct synthetic partial multi-label datasets with known ground-truth labels, then add false candidate labels at several noise levels (say 10%, 30%, 50%, and 70%) and run PML-FSMIR on each; if its ranking loss or F1 degrades faster than a baseline that ignores label structure, or if the reconstructed label matrix $T$ disagrees with the known clean labels, the noise-resistance premise fails. The paper only tests one fixed 20% noise setting and never compares the reconstructed labels against a clean-label ground truth, so this check would settle the claim.","tokens_in":13933,"feed_emoji":"🎯","tokens_out":9212,"duration_ms":87043,"temperature":0.7,"pith_summary":"The paper proposes a feature-selection method for partial multi-label learning, the setting where each example carries a candidate label set and at least one label is correct while the rest may be noise. The method, PML-FSMIR, uses mutual information between labels to rebuild the noisy label matrix, learns feature weights without collapsing the data to a low-rank space, and finally reweights features toward labels that are strongly connected to many other labels. The paper reports that on eight benchmark datasets and five evaluation metrics this procedure ranks first in every combination except Coverage on CHD49 and Average Precision on LLOGF, where it ranks second. A sympathetic reader would take the paper's claim to be that label-relationship structure is stable enough under label noise to serve as the foundation for both label disambiguation and reliable feature selection.","feed_headline":"A three-stage method wins nearly all noisy multi-label benchmarks","feed_subtitle":"PML-FSMIR cleans candidate labels by mutual information, then reweights features toward well-connected labels.","key_machinery":"The central objects are the two mutual-information matrices: $Z_{ij} = I(y_{:i}, y_{:j})$ computed from the observed candidate labels, and $Z'_{ij} = I(T_{:i}, T_{:j})$ computed from the reconstructed label matrix. The first matrix enters $T = (Y Z) \\circ \\mathrm{sign}(Y)$, reweighting each candidate label by the summed mutual information it shares with the other candidate labels in that sample; this is the noise-resistance mechanism, since an isolated candidate label with weak associations is downweighted as likely noise. The second matrix supplies the graph-Laplacian regularizer during weight learning and performs the final update $W \\leftarrow W Z'$, which shifts weight toward features that identify labels with many strong connections. The reformed low-rank assumption keeps $X$ at its original dimension by using $UV$ as a denoised reconstruction of $X$ rather than as a replacement for it.","core_discovery":"PML-FSMIR claims that two collective properties of labels — noise-resistant label relationships and label connectivity — are sufficient for feature selection in partial multi-label learning without relying on the low-rank assumption used by earlier PML methods. The paper constructs a label-reconstruction matrix $T = (Y Z) \\circ \\mathrm{sign}(Y)$ from the candidate-label matrix $Y$ and the pairwise label mutual-information matrix $Z$, so a candidate label is treated as reliable when the other candidate labels in the same sample are strongly associated with it. The learning objective then combines $\\|UVW - T\\|_F^2$, $\\alpha\\|X - UV\\|_F^2$, a manifold regularizer built from the mutual-information graph of $T$, and a $\\gamma\\|W\\|_{2,1}$ term for feature selection, with $U$ and $V$ reconstructing $X$ while keeping its full dimension. In the final stage the learned weight matrix is replaced by $W Z'$, which magnifies weights for features that discriminate representative, well-connected labels. The paper argues that this ordering of operations removes label noise before learning, avoids information loss from low-rank projection, and concentrates the selected features on labels that are important by connectivity.","pith_inferences":["If the co-occurrence-based reconstruction is as effective as reported, the natural next test is to vary the noise level instead of keeping it fixed at 20%, since the reconstruction's premise is that label co-occurrence statistics survive moderate noise.","The same mutual-information reweighting could transplant directly into deep partial multi-label classifiers, where the label graph is often available but label noise is handled by separate modules.","The weight-reconstruction rule implies a testable ranking hypothesis: features with high weight on high-degree labels should transfer to better detection of rare positive labels, which could be checked on a dataset with explicit label-degree annotations.","The reported F1 improvements are large in absolute terms, so a paired significance test across the ten-fold cross-validation splits would clarify whether the gap is stable rather than driven by a few folds."],"forward_implications":["Ranked feature subsets produced by the method should improve SVM classification on partial multi-label data across audio, music, text, image, chemistry, biology, and medicine benchmarks.","The third-stage weight reconstruction is what makes positive-label identification possible: without it the paper reports that the method identifies no positive labels on six of eight datasets, and with it the method identifies positive labels on all eight.","Because the method keeps the original feature dimension rather than replacing $X$ by a low-rank factor, it should retain high-order structural information that low-rank PML methods discard.","The method's ranking behavior is stable to the three regularization parameters $\\alpha$, $\\beta$, and $\\gamma$ across the tested range from 0.001 to 1000."],"supporting_citations":[{"why":"Defines the partial multi-label learning problem and supplies the PML-LC and PML-FP baselines the method must beat.","marker":"[Xie and Huang, 2018]"},{"why":"Introduces the low-rank and sparse decomposition approach for PML that the paper critiques and reforms.","marker":"[Sun et al., 2019]"},{"why":"Supplies the FPML baseline, a feature-induced partial multi-label learning method used for comparison.","marker":"[Yu et al., 2018]"},{"why":"Supplies the PARTICAL-MAP and PARTICAL-VLS baselines that elicit credible labels from the candidate set.","marker":"[Zhang and Fang, 2020]"},{"why":"Provides the definition of mutual information that underlies both label-reconstruction stages.","marker":"[Shannon, 1948]"},{"why":"Supplies the $\\ell_{2,1}$-norm regularization used to make the learned weight matrix select features jointly.","marker":"[Nie et al., 2010]"},{"why":"Supplies the only prior partial multi-label feature-selection method, PML-FSSO, used as a direct baseline.","marker":"[Hao et al., 2023]"},{"why":"Supplies the MIFS multi-label informed feature-selection baseline.","marker":"[Jian et al., 2016b]"},{"why":"Supplies the DRMFS dual-graph regularized multi-label feature-selection baseline.","marker":"[Hu et al., 2020a]"}],"fun_headline_variants":["Noise-resistant label reconstruction boosts partial multi-label learning","Three-stage method cleans labels then selects features, beats baselines","PML without low-rank: feature selection via label connectivity","Label relationships and connectivity win noisy multi-label benchmarks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the label-to-label association statistics computed from the observed candidate labels are still trustworthy, so that a candidate label surrounded by associated candidate labels is likely correct and an isolated one is likely noise; if the label noise is strong enough to distort those statistics, the reconstruction would amplify the noise instead of removing it.","fun_headline_variants_meta":{"raw":{"variants":["Noise-resistant label reconstruction boosts partial multi-label learning","Three-stage method cleans labels then selects features, beats baselines","PML without low-rank: feature selection via label connectivity","Label relationships and connectivity win noisy multi-label benchmarks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000453,"raw_usage":{"total_tokens":2304,"prompt_tokens":997,"completion_tokens":1307,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":613,"completion_tokens_details":{"reasoning_tokens":1242}},"tokens_in":613,"tokens_out":1307,"duration_ms":10934,"temperature":1.0,"reasoning_tokens":1242,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:36:34.817528+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct synthetic partial multi-label datasets with known ground-truth labels, then add false candidate labels at several noise levels (say 10%, 30%, 50%, and 70%) and run PML-FSMIR on each; if its ranking loss or F1 degrades faster than a baseline that ignores label structure, or if the reconstructed label matrix $T$ disagrees with the known clean labels, the noise-resistance premise fails. The paper only tests one fixed 20% noise setting and never compares the reconstructed labels against a clean-label ground truth, so this check would settle the claim.","supporting_citations":[{"cited_title":"Partial multi-label learning","cited_arxiv_id":null,"evidence_quote":"Defines the partial multi-label learning problem and supplies the PML-LC and PML-FP baselines the method must beat."},{"cited_title":"Partial multi-label learning by low-rank and sparse decomposition","cited_arxiv_id":null,"evidence_quote":"Introduces the low-rank and sparse decomposition approach for PML that the paper critiques and reforms."},{"cited_title":"Partial multi-label learning via credible label elicita- tion","cited_arxiv_id":null,"evidence_quote":"Supplies the PARTICAL-MAP and PARTICAL-VLS baselines that elicit credible labels from the candidate set."},{"cited_title":"A mathematical theory of communication","cited_arxiv_id":null,"evidence_quote":"Provides the definition of mutual information that underlies both label-reconstruction stages."},{"cited_title":"Efficient and robust feature selection via joint l2, 1-norms minimization","cited_arxiv_id":null,"evidence_quote":"Supplies the $\\ell_{2,1}$-norm regularization used to make the learned weight matrix select features jointly."},{"cited_title":"Partial multi-label feature selection via subspace optimiza- tion","cited_arxiv_id":null,"evidence_quote":"Supplies the only prior partial multi-label feature-selection method, PML-FSSO, used as a direct baseline."}],"review_version":1}