{"id":"d7617ea1-b8fe-4be0-b9aa-7f1f1875aa8a","arxiv_id":"2411.17128","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":2.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Replacing the fixed threshold 2 in SFFSVM with a per-dataset tuned location parameter yields small average gains on imbalanced KEEL datasets.","lead":"This paper modifies a fuzzy support vector machine by replacing a fixed threshold with a tunable location parameter, and reports modest F1, MCC, and AUC-PR gains over prior SVM-based baselines on imbalanced KEEL datasets. The generalist takeaway is that tuning one threshold can improve imbalance metrics, but the reported gains are small and lack error bars.","discovery_kind":"incremental","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Per-dataset tuning of a on a 10-point grid, rather than the unproved geometric 'shift right' mechanism, may explain the small average gains; the 'consistently higher' claim needs nested evaluation and significance tests.","rationale":"The reader identified the geometric 'shift right' assumption as weakest; I agree it is unproven, but on reading the experimental section the more operational weakness is that the only difference from SFFSVM is the per-dataset grid-selected threshold a, with a=2 recovering SFFSVM. Thus the reported improvements are expected to be non-negative on the tuning folds by construction, and the main-text tables do not establish that they are significant on held-out data. The Friedman/Nemenyi tests are relegated to the supplementary and are not accompanied by per-dataset evidence in the paper. This does not require rejecting the method: the code is available, the change is simple, and the proposed check is computational. However, the evaluation must be strengthened before the 'consistently higher' claim is accepted. My recommendation therefore matches the reader's CONDITIONAL verdict.","tokens_in":13846,"tokens_out":6826,"duration_ms":62521,"concrete_test":"Run all KEEL datasets under nested CV: for each dataset, split into 80/20; on the 80% training part, use 5-fold CV to pick a from [1.1:0.1:2] maximizing the target metric; retrain on the full 80% with that a and evaluate on the 20% test part; repeat for 10 independent splits. Report paired ISFFSVM-SFFSVM differences per dataset and run Wilcoxon signed-rank tests at a=0.05 for F1, MCC, and AUC-PR. Additionally, run ISFFSVM with one fixed a equal to the median selected value across datasets; if the fixed-a version no longer beats SFFSVM significantly, the gains are attributable to per-dataset threshold selection rather than the location-parameter mechanism.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim rests on two linked assumptions. First, Section III.A asserts without proof or margin argument that lowering the fuzzy memberships of majority points with slack factors in [a,2) in Eq. (9) moves the DEC hyperplane Mdec 'right' and saves correctly classified minority points; the only evidence is a single 2D moons example (Fig. 3). Second, the empirical support is confounded by model selection: ISFFSVM tunes a over {1.1,1.2,...,2.0} per dataset, and since a=2 reduces ISFFSVM to SFFSVM, the proposed method is guaranteed to be at least as good as SFFSVM on the tuning criterion. The reported margins are small (e.g., F1 +1.01 on low-IR and +1.73 on high-IR, Tables I/II) and appear only as averages, without per-dataset paired differences, error bars, or the significance tests that are mentioned in Section IV.B but deferred to the supplementary. If the gains are not robust under nested cross-validation and paired testing, they reflect threshold picking over the extra hyperparameter rather than the claimed geometric improvement.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an improved slack-factor-based fuzzy support vector machine (ISFFSVM) for class-imbalanced learning. The novelty is a location parameter a in the fuzzy membership function for majority-class samples, which replaces the fixed threshold of 2 used in the baseline SFFSVM. The authors argue that this change reduces the risk of misclassifying correctly classified minority samples when shifting the DEC hyperplane. The method is evaluated on KEEL datasets grouped by imbalance ratio (IR < 10 and IR >= 10), using average F1-score, MCC, and AUC-PR, and additionally on a schizophrenia dataset. The reported results show small average improvements over baseline classifiers, and the paper includes a sensitivity analysis of the location parameter.","tokens_in":14119,"tokens_out":4565,"duration_ms":38837,"significance":"If the claimed improvements are statistically robust, this is a modest but useful extension of fuzzy SVM methods for class imbalance. The paper provides a simple algorithmic change, releases code, and uses standard public benchmarks, which are positive features. However, the theoretical justification is heuristic and the empirical evidence in the main text is incomplete, because only averages are reported and the significance tests are deferred to a supplementary that is not available for review. The work is likely to be of interest to the fuzzy SVM community, but the central claim of consistent improvement requires stronger support.","major_comments":[{"comment":"The theoretical justification for the central mechanism is an assertion rather than a proof. The claim that assigning lower membership values to majority samples with slack factors in [a,2) better positions the DEC hyperplane Mdec when shifted to the right and reduces misclassification of correctly classified minority samples is argued entirely through the 2D example in Fig. 2 and the moons dataset in Fig. 3. No margin argument, bound, or high-dimensional validation is provided. This is load-bearing because if the mechanism does not hold generally, the reported gains reduce to per-dataset threshold selection. The authors should either provide a formal argument relating the membership weights to the optimal hyperplane or substantially soften the claim to say that a can be tuned to improve performance.","section":"Section III.A"},{"comment":"The empirical support for the 'consistently achieves higher' claim is incomplete. Tables I and II report only average F1, MCC, and AUC-PR over datasets, with no standard deviations, per-dataset paired differences, confidence intervals, or effect sizes. The paper mentions a Friedman test and Nemenyi post hoc test in Section IV.B but defers details to a supplementary that is not included in this manuscript. Furthermore, because ISFFSVM tunes a over the grid {1.1,...,2.0} per dataset and a=2 recovers SFFSVM, ISFFSVM is guaranteed to be at least as good as SFFSVM on the validation set. The reported average gains (e.g., F1 +1.01 on low IR and +1.73 on high IR) are therefore consistent with threshold picking rather than the proposed geometric mechanism. The authors should report the full statistical analysis in the main text, including paired tests and confidence intervals, and ideally a nested cross-validation where a is selected on the training fold only.","section":"Section IV.B, Tables I and II"},{"comment":"The sensitivity analysis demonstrates that the optimal a varies widely across datasets (e.g., 1.1 for Haberman, 1.7 for Ecoli1) and that performance is highly sensitive to a, with steep declines when deviating from the best value (as in Haberman). This undercuts the paper's framing of a as a novel parameter that significantly advances the model in a principled way; it suggests the gains arise from per-dataset hyperparameter optimization. The authors should acknowledge this limitation explicitly and discuss the cost of tuning a relative to the small average improvements.","section":"Section IV.D"}],"minor_comments":[{"comment":"The typesetting of the objective functions is confusing: the summation indices such as \\sum_{x \\in X^+} are placed after the weight terms, making the formulas hard to read; they should appear before the terms they multiply.","section":"Equations (1) and (7)"},{"comment":"The term \"right\" is used informally to describe the direction of the DEC hyperplane shift; a more precise definition, for example in terms of the change in the bias b or the normal vector w, would improve rigor and reproducibility.","section":"Section III.A"},{"comment":"The abstract claims ISFFSVM \"consistently achieves higher F1-scores\" but the main text only reports averages without error bars or significance tests; the language should be tempered until statistical validation is presented.","section":"Abstract and Section IV.B"},{"comment":"The x-axis tick labels in Fig. 4 appear to be truncated or incorrectly placed, showing values up to 1.9 and then \"255\" instead of the full range from 1.1 to 2.0; the figure should be corrected.","section":"Figure 4"},{"comment":"The supplementary tables (S.V–S.X) and the statistical analysis section S.I.A are referenced but not included in the manuscript; for a self-contained review, either include them or summarize the key results in the main text.","section":"Supplementary"},{"comment":"The \"Error Analysis\" subsection is qualitative and largely repeats the mechanism described in Section III.A; it does not provide quantitative error bounds or a formal analysis of false positives and false negatives, so its length could be reduced.","section":"Section III.B"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of the journal, but the main claims hinge on the supplementary statistical analysis and on the interpretation of the tuning procedure. The fact that a=2 recovers SFFSVM means that the comparison to SFFSVM is not a fair test of the proposed mechanism; the authors should report results where the baseline SFFSVM is given the same tuning budget, or explicitly acknowledge the nested-selection issue. The lack of standard deviations and per-dataset comparisons in the main text is a serious concern for a paper whose title promises a 'novel' approach with 'consistently' better performance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a one-line modification to SFFSVM—replacing the fixed threshold 2 in the membership function with a tunable location parameter a—and the paper knows it, stating that a=2 reduces the method to SFFSVM. What's genuinely useful: the authors are transparent about that reduction, they release code, and they run a broad set of KEEL comparisons against many baselines. The average gains are small but in a sensible direction: about one F1 point on low IR, about 1.7 on high IR, with similar movements in MCC and AUC-PR. That is the most honest reading of the contribution: a careful practitioner could adopt ISFFSVM as a drop-in improvement at the cost of one extra hyperparameter.\n\nWhere the paper oversells: the abstract says 'consistently achieves higher F1-scores, MCC, and AUC-PR compared to baseline classifiers,' but the main text only shows averages, with no standard deviations or confidence intervals. The significance tests mentioned in Section IV.B are deferred to the supplementary, which is fine if they are actually there, but the main text should say what they show. More importantly, the comparison is not apples-to-apples: ISFFSVM tunes a per dataset over [1.1, 2], while SFFSVM uses the fixed value 2. Since a=2 is in the grid, ISFFSVM is guaranteed to match SFFSVM at worst, and the reported gains could be driven by hyperparameter search rather than the geometric 'shift right' mechanism claimed in Section III.A. That mechanism is illustrated on a single 2D moons dataset; there is no margin argument or high-dimensional validation. So the 'consistently higher' claim is too strong as stated.\n\nThe theoretical justification section is really a verbal explanation, not theory—there is no theorem or bound. That is a minor overclaim if the paper is treated as an empirical incremental study, which I think is the right frame.\n\nVerdict: the paper is a legitimate incremental empirical contribution. It deserves a serious referee, but it needs a revision that reports per-dataset paired differences with variance, moves the significance testing into the main text or states its absence honestly, and ideally adds a comparison where SFFSVM also gets a small hyperparameter search over the same grid restricted to a=2, to isolate the effect of the tunable threshold. If those additions hold up, the method is a reasonable contribution to the fuzzy SVM imbalance niche. I'd cite it if I worked in that niche, and I'd bring it to a reading group focused on fuzzy SVM methods.","headline":"Tunable-threshold SFFSVM with honest self-awareness and code; the small gains are plausible but the 'consistently higher' claim needs variance estimates and a fairer baseline search before it can be taken at face value.","tokens_in":14656,"tokens_out":2732,"would_cite":false,"duration_ms":23991,"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":"Replacing a fixed threshold with a tunable location parameter in the fuzzy membership function reduces minority-class misclassification on imbalanced datasets.","keywords":["class imbalance","fuzzy support vector machine","slack factor","location parameter","fuzzy membership","different error cost","imbalanced classification","AUC-PR"],"falsifier":"Run ISFFSVM across many higher-dimensional imbalanced datasets and check whether the optimal $a$ is systematically less than 2 and whether the number of correctly classified minority samples actually increases as $a$ decreases; if the best $a$ clusters at 2 or performance gains vanish when features increase, the location-parameter mechanism is not general. A direct check on real datasets is to record the false-negative count on the minority class for each $a$ in the range and test whether it decreases monotonically as claimed by the geometric story.","tokens_in":76,"feed_emoji":"⚖️","tokens_out":5422,"duration_ms":85244,"temperature":0.7,"pith_summary":"This paper claims that a one-line change to the fuzzy membership rule of the slack-factor-based fuzzy SVM (SFFSVM) improves performance on class-imbalanced datasets. In SFFSVM, majority-class samples with slack factor below 2 receive full fuzzy membership; the proposed ISFFSVM replaces the fixed cutoff 2 with a tunable location parameter a. By assigning lower membership to majority samples whose slack factor lies between a and 2, the decision hyperplane from the different error cost (DEC) model is kept from shifting too far, which protects correctly classified minority samples from being misclassified. If correct, this gives a simple algorithmic-level remedy for imbalance that adds no computational cost beyond tuning one parameter, and the paper reports higher average F1-score, MCC, and AUC-PR on real-world datasets.","feed_headline":"Tunable threshold boosts fuzzy SVM on imbalanced data","feed_subtitle":"A simple parameter change in SFFSVM reduces minority misclassification and lifts F1, MCC, and AUC-PR across benchmark datasets.","key_machinery":"The load-bearing object is the location parameter $a$ in Eq. (9), a tunable cutoff in the fuzzy membership function for the majority class. In the prior SFFSVM, majority points with slack factor $\\xi < 2$ all receive membership 1; ISFFSVM grants full membership only to points with $\\xi < a$ and exponentially decays membership for $\\xi \\geq a$. This controls how far the DEC hyperplane $M_{dec}$ can shift rightward when the model retrains with these memberships: points with slack factors between $a$ and 2, which the old rule would weight at 1, now receive lower weight, so the boundary does not push past correctly classified minority points. The machinery also includes the DEC model of Eq. (7), which supplies the slack factors, and the two-stage training loop of Algorithm 1 that computes memberships from DEC slacks and then solves Eq. (10).","core_discovery":"The central discovery is that the fixed value 2 used in SFFSVM's membership function for majority-class samples is not always the right threshold for shifting the DEC hyperplane, and that making this threshold a tunable location parameter a, with a chosen from 1.1 to 2, reduces the risk of misclassifying minority samples. Equations (8)-(9) set minority memberships as before, while majority memberships are 1 for slack factor $\\xi < a$ and $e^{-\\mu \\xi}$ for $\\xi \\geq a$; for $a=2$ the model reduces to SFFSVM. The paper asserts, with 2D visual illustration and experiments on real-world imbalanced datasets, that this selective lower weighting of majority points near the boundary positions the shifted DEC hyperplane so that more minority samples are correctly classified, yielding consistently higher F1-score, MCC, and AUC-PR than baseline classifiers.","pith_inferences":["The geometric story is only demonstrated on two-dimensional synthetic data; in higher dimensions the rightward shift of the DEC hyperplane may not be monotonic in $a$, so the reported gains could partly reflect per-dataset threshold selection rather than a universal mechanism.","A testable extension is an adaptive schedule that adjusts $a$ during training based on the current false-negative rate, removing the grid search.","The same thresholding idea could transfer to other fuzzy SVM variants or cost-sensitive classifiers, where membership cutoffs are currently fixed.","Because average gains over SFFSVM are roughly one percentage point, the practical value depends on variance across datasets and on the statistical tests reported in the supplementary material."],"forward_implications":["For $a=2$, ISFFSVM reduces exactly to SFFSVM, so the proposal is a strict generalization that can only help when a different threshold fits the data better.","On both low-imbalance and high-imbalance public benchmark datasets, ISFFSVM reports higher average F1-score, MCC, and AUC-PR than SFFSVM and other baselines.","The computational complexity stays $O(dN^2)$, the same as SFFSVM, with the only extra cost being grid-search tuning of $a$.","On a schizophrenia diagnosis dataset, ISFFSVM improves slightly over SFFSVM on F1-score, MCC, and AUC-PR.","The model is most beneficial under severe class disparity, where preserving minority samples matters most."],"supporting_citations":[{"why":"Defines the slack-factor-based fuzzy SVM and the fixed threshold 2 that ISFFSVM generalizes with the location parameter $a$.","marker":"[24]"},{"why":"Provides the different error cost (DEC) model whose hyperplane $M_{dec}$ is shifted in the proposed membership scheme.","marker":"[21]"},{"why":"FSVM-CIL, the prior fuzzy class-imbalance method whose misclassification issue motivates the slack-factor and location-parameter adjustments.","marker":"[23]"},{"why":"Supplies the real-world imbalanced datasets used to compare ISFFSVM against baseline classifiers.","marker":"[32]"},{"why":"Introduces fuzzy memberships to SVM, the background that SFFSVM and ISFFSVM build on.","marker":"[22]"}],"fun_headline_variants":["Location parameter tunes fuzzy SVM for imbalanced data","Improved slack-factor fuzzy SVM with adaptive threshold","Tunable threshold lifts minority recall in fuzzy SVM","New parameter helps fuzzy SVM handle class imbalance","ISFFSVM: better fuzzy SVM via location parameter"],"cache_read_input_tokens":16768,"weakest_assumption_plain":"The argument depends on the geometric assumption that lowering fuzzy memberships for majority samples with slack factors in $[a,2)$ shifts the decision boundary rightward just enough to save correctly classified minority points without harming them; this is illustrated only on a two-dimensional synthetic dataset and asserted without proof in higher dimensions.","fun_headline_variants_meta":{"raw":{"variants":["Location parameter tunes fuzzy SVM for imbalanced data","Improved slack-factor fuzzy SVM with adaptive threshold","Tunable threshold lifts minority recall in fuzzy SVM","New parameter helps fuzzy SVM handle class imbalance","ISFFSVM: better fuzzy SVM via location parameter"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000176,"raw_usage":{"total_tokens":1335,"prompt_tokens":1037,"completion_tokens":298,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":653,"completion_tokens_details":{"reasoning_tokens":226}},"tokens_in":653,"tokens_out":298,"duration_ms":3107,"temperature":1.0,"reasoning_tokens":226,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:28:52.182658+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run ISFFSVM across many higher-dimensional imbalanced datasets and check whether the optimal $a$ is systematically less than 2 and whether the number of correctly classified minority samples actually increases as $a$ decreases; if the best $a$ clusters at 2 or performance gains vanish when features increase, the location-parameter mechanism is not general. A direct check on real datasets is to record the false-negative count on the minority class for each $a$ in the range and test whether it decreases monotonically as claimed by the geometric story.","supporting_citations":[{"cited_title":"Slack-factor-based fuzzy support vector machine for class imbalance problems,","cited_arxiv_id":null,"evidence_quote":"Defines the slack-factor-based fuzzy SVM and the fixed threshold 2 that ISFFSVM generalizes with the location parameter $a$."},{"cited_title":"Con- trolling the sensitivity of support vector machines,","cited_arxiv_id":null,"evidence_quote":"Provides the different error cost (DEC) model whose hyperplane $M_{dec}$ is shifted in the proposed membership scheme."},{"cited_title":"FSVM-CIL: fuzzy support vector machines for class imbalance learning,","cited_arxiv_id":null,"evidence_quote":"FSVM-CIL, the prior fuzzy class-imbalance method whose misclassification issue motivates the slack-factor and location-parameter adjustments."},{"cited_title":"KEEL data-mining software tool: Data set repository, integra- tion of algorithms and experimental analysis framework,","cited_arxiv_id":null,"evidence_quote":"Supplies the real-world imbalanced datasets used to compare ISFFSVM against baseline classifiers."},{"cited_title":"Fuzzy support vector ma- chines,","cited_arxiv_id":null,"evidence_quote":"Introduces fuzzy memberships to SVM, the background that SFFSVM and ISFFSVM build on."}],"review_version":1}