{"id":"157a33a4-99dc-48fa-944e-07091d01834b","arxiv_id":"2504.16268","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"The paper applies existing opposition-based reflection to augment KNN training data, but the reported results do not support the claimed consistent improvements.","lead":"This paper proposes reflecting training data across feature ranges, an opposition-based learning idea, to boost KNN classifiers. The reported 26-dataset evaluation does not show consistent gains, and the two most distinctive variants are missing from the results.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claim rests on an unspecified clean-evaluation protocol: Section 5 computes Z-score statistics, mutual-information feature selection, and OBL min/max bounds before any train/test split is described, so if these use the full dataset, every reported OBL gain may be a data-leakage artifact.","rationale":"The reader's weakest assumption—that all preprocessing statistics are computed from training folds only—is precisely the condition on which the central claim rests. The paper never states this condition and describes its pipeline (Section 5) in a way that suggests preprocessing is applied before model training, with cross-validation only mentioned later as an evaluation scheme (Section 6.1). This is the most load-bearing concern because it would invalidate every reported number, making the abstract's 'consistently outperform' claim unsupported even before considering other weaknesses. The reader's additional points (tables contradicting the consistency claim, missing class-wise/local variants, absent Friedman figures) are secondary: even if those were fixed, the leakage ambiguity would remain. Since my analysis confirms and sharpens the reader's concern, the verdict of REJECT remains appropriate; I mark UNCHANGED because my critique does not move the verdict but reinforces it.","tokens_in":13641,"tokens_out":5830,"duration_ms":55073,"concrete_test":"Re-run the full Section 5 pipeline with an explicit 5-fold cross-validation where every preprocessing step (Z-score parameters, mutual-information feature selection, and OBL global/class-wise/local min/max bounds, plus post-augmentation Z-score normalization) is recomputed on the training fold only, then evaluated on the held-out fold. Compare the resulting OBLKNN and WOBLKNN accuracies to Tables 2 and 3. If the OBL advantage shrinks or reverses on datasets where OBL currently appears superior (e.g., datasets 15, 16, 23, 24), leakage is confirmed and the central claim falls.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim—that OBL-enhanced classifiers consistently outperform basic KNN—depends entirely on the reported accuracy and F1 numbers. Those numbers are only meaningful if all preprocessing statistics are computed on training folds exclusively. Section 5's pipeline (Z-score normalization, optional mutual-information feature selection, OBL transformation using global/class-wise/local min-max bounds, and post-augmentation Z-score normalization) is presented as a linear sequence ending in model training, with no mention of cross-validation or fold boundaries. Section 6.1 states only that a 5-fold CV was used and results averaged over 30 runs; it does not state that the pipeline was refit inside each fold. If any statistic—the Z-score mean/standard deviation, the selected feature subset, or the OBL bounds (a_k, b_k, a_{c,k}, b_{c,k}, a_{i,k}, b_{i,k})—is computed on the full dataset before splitting, then the transformed training data encodes information from the held-out test fold, and the KNN accuracies in Tables 2 and 3 are optimistically biased. This would invalidate the central claim regardless of how the tables are read. The concern is not merely hypothetical: the manuscript nowhere provides the code or an explicit statement of fold-aware preprocessing, and the described pipeline in Section 5 conspicuously omits any split step.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes three opposition-based learning (OBL) transformations—Global OBL, Class-Wise OBL, and Localized Class-Wise OBL—as a data-augmentation preprocessing step for KNN and other classifiers. It claims that OBL-enhanced classifiers consistently outperform plain KNN across 26 datasets, with evaluation by 5-fold cross-validation averaged over 30 runs and reported accuracy, F1, and runtime. The manuscript describes the OBL schemes, provides a preprocessing pipeline, and presents tables comparing KNN, WKNN, OBLKNN, and WOBLKNN, plus Friedman test figures for accuracy, F1, and runtime.","tokens_in":14000,"tokens_out":4249,"duration_ms":36829,"significance":"If the central claim were sound, the contribution would be practically useful: a simple, geometry-aware reflection of training samples could improve KNN in high-dimensional or sparse settings with negligible overhead. The paper has several strengths in presentation: the OBL variants are clearly defined in Section 5, the dataset collection is sizable and includes many high-dimensional problems, and the authors report multiple metrics and average over 30 runs. However, the central claim is not supported by the paper's own tables, the evaluation protocol is under-specified, and the proposed class-wise and localized variants are absent from the reported experimental results. These issues are load-bearing for the paper's contribution, so the current manuscript does not establish its advertised finding.","major_comments":[{"comment":"The experimental protocol does not specify that all preprocessing statistics are computed on the training folds only. The pipeline in Section 5 lists Z-score normalization, mutual-information feature selection, OBL bounds (a_k, b_k, a_{c,k}, b_{c,k}, a_{i,k}, b_{i,k}), and post-augmentation normalization as a linear sequence ending in model training, with no mention of a train/test split. Section 6.1 states only that a 5-fold CV was used and results were averaged over 30 runs, not that the pipeline was refit inside each fold. If any of these statistics are computed on the full dataset before splitting, the accuracy and F1 values in Tables 2 and 3 are optimistically biased and the central claim is invalidated. This is the most serious issue because it affects every reported result.","section":"Section 5 and Section 6.1"},{"comment":"The claim that OBL-enhanced classifiers 'consistently outperform' basic KNN is contradicted by the paper's own numbers. In Table 2, OBLKNN accuracy is lower than KNN accuracy on 10 of the 26 datasets, including dataset 1 (0.9250 vs. 0.9256), dataset 3 (0.9247 vs. 0.9300), dataset 4 (0.8494 vs. 0.8497), dataset 9 (0.7442 vs. 0.7448), dataset 13 (0.9804 vs. 0.9807), dataset 17 (0.9892 vs. 0.9901), dataset 19 (0.6950 vs. 0.6952), dataset 20 (0.9478 vs. 0.9486), dataset 21 (0.9252 vs. 0.9253), and dataset 25 (0.9591 vs. 0.9607). A pattern of mixed and often tiny differences does not support the abstract's 'consistently outperform' wording.","section":"Abstract and Table 2"},{"comment":"The paper's proposed class-wise and localized variants are central to the contribution, but no experimental results are reported for them. Section 6 states that 'Class-Wise Opposition-Based Learning is the most performed,' and Section 7.1 says the evaluation covers OBLKNN-CW, LOBLKNN-CW, WOBLKNN-CW, and WLOBLKNN-CW. However, Tables 2-4 contain only KNN, WKNN, OBLKNN, and WOBLKNN. The Friedman test figures (Figures 3-5) are referenced but their contents, p-values, and post-hoc comparisons are not reported in the text. The selection of a best variant is therefore unsupported by any presented evidence.","section":"Section 7.1 and Tables 2-4"},{"comment":"There is an internal inconsistency about whether OBL generates additional samples or replaces the original data. The surrounding text says the scheme 'enriches the training set' and describes an 'OBL-augmented' dataset, but one pipeline bullet says the transformation produces 'oppositional samples that replace the original dataset,' and another bullet says training is done on the 'OBL-augmented and normalized dataset.' If opposite samples replace rather than augment, the method is not data augmentation as claimed, and the reported comparisons are not testing the stated approach. The authors need to state unambiguously which dataset is used for training.","section":"Section 5, pipeline bullets"},{"comment":"No statistical uncertainty is reported for any of the accuracy or F1 values: there are no standard deviations, confidence intervals, p-values, or effect sizes. The manuscript claims Friedman tests were conducted, but the text does not report the test statistics or associated p-values, and the figures referenced as evidence are not reproduced in the manuscript body. Given that many pairwise differences in Table 2 are on the order of 0.001 or less, the claim of statistically validated improvement is not currently testable.","section":"Section 7.1 and Figures 3-5"}],"minor_comments":[{"comment":"The phrase 'adversarial-based learning (OBL)' near the end of Section 1 should read 'opposition-based learning (OBL)'.","section":"Introduction"},{"comment":"Equation numbers are not unique: Equation (1) appears for min-max scaling in Section 3.1 and again for generalized OBL in Section 4.1; similar duplicate numbering occurs for other equations. Please renumber equations sequentially throughout the manuscript.","section":"Equation numbering"},{"comment":"The reference list contains inconsistencies and editorial notes, including duplicate entries for [43], a reference marked as a duplicate of [42], and a mismatched in-text citation for OBL: the text cites Tizhoosh's OBL paper as [7], but the reference list entry [7] is the ADASYN paper, with Tizhoosh at [8].","section":"References"},{"comment":"The tables do not include dataset names or a key row label; each row is identified only by a number. Adding abbreviated dataset names would make it much easier for readers to check individual results.","section":"Tables 2-4"},{"comment":"For the Semeion dataset, the class count is listed as 2, but Semeion is usually a 10-class handwritten digit dataset; if the data were binarized, this should be stated explicitly.","section":"Table 1"}],"recommendation":"reject","confidential_remarks":"I concur with the reader's assessment and recommend rejection. The most defensible reading is that the central claim is contradicted by the manuscript's own results and that the evaluation protocol is under-specified to the point where leakage cannot be ruled out. The paper has a reasonable idea, but as written it does not establish the claimed improvement. If the authors were to refit all preprocessing inside cross-validation folds, report the class-wise and localized variants they claim to evaluate, and provide uncertainty quantification, a substantially revised version could be reconsidered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi,\n\nThe idea here is simple: use the classical OBL reflection x* = a+b−x as a data augmentation step before training KNN, with three variants (global, class-wise, local class-wise). Applying OBL to classification preprocessing is new as far as I know. The paper runs 26 datasets, mostly high-dimensional, which is a reasonable test bed.\n\nWhat works: the variants are clearly defined. The class-wise and local versions are a modest extension of the global formula, but they are not identical to it; the local version uses nearest-neighbor bounds, which is a bit more than just changing interval endpoints. The experimental design (5-fold CV, 30 runs) is standard.\n\nThe problems outweigh the credit. The tables don't include the two variants the paper says are the best. Table 2 only reports KNN, WKNN, OBLKNN, and WOBLKNN. The class-wise and local results are absent, so the central claim about those variants is unverifiable. The abstract says 'consistently outperform,' but Table 2 shows OBLKNN below plain KNN on 10 of 26 datasets (e.g., dataset 1: 0.925 vs 0.9256; dataset 3: 0.9247 vs 0.93). That's not consistent. No standard deviations, p-values, or effect sizes are reported; the Friedman test figures are missing. Most concerning, the pipeline in Section 5 computes Z-score statistics, mutual-information feature selection, and OBL bounds before mentioning any train/test split. If those are computed on the whole dataset, every reported gain is potentially a leakage artifact. The stress-test note is correct: the paper never states that preprocessing is refit inside each fold.\n\nThe manuscript is also unfinished. The conclusion section is just the word 'Conclusion.' The references contain editorial notes and duplicated entries. That signals the submission is not ready.\n\nWho gets value from this? Someone curious about simple augmentation tricks for KNN on tabular data might implement the idea themselves, but the paper doesn't support its claims. I'd desk reject it. If the editor wants to be generous, request a complete manuscript with fold-aware preprocessing, full tables for all variants, error bars, and code. As it stands, it's not ready for serious peer review.","headline":"A simple OBL reflection trick for KNN augmentation, undermined by incomplete reporting, missing results, and a likely data-leakage problem.","tokens_in":14442,"tokens_out":3622,"would_cite":false,"duration_ms":31613,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Opposition-based mirror images boost KNN accuracy across 26 datasets.","keywords":["Opposition-Based Learning","data augmentation","K-nearest neighbors","classification","high-dimensional data","data transformation","class-wise OBL","feature selection"],"falsifier":"Run the same 26-dataset protocol with a strict fold-wise pipeline in which Z-score normalization, mutual-information feature selection, and OBL min/max bounds are recomputed inside each training fold only, then compare OBL-KNN against plain KNN; if the accuracy advantage disappears or shrinks to statistical noise, the original gains came from leakage of test-set information into preprocessing.","tokens_in":13460,"feed_emoji":"🔄","tokens_out":8206,"duration_ms":68378,"temperature":0.7,"pith_summary":"The paper proposes using Opposition-Based Learning (OBL)—a reflection operation from optimization—to create synthetic 'opposite' training samples for classification. It tests three variants, with bounds taken from the whole dataset, from each class, or from local same-class neighborhoods, and combines them with K-nearest neighbors (KNN). Across 26 datasets, the reported accuracy and F1-scores for OBL-augmented KNN consistently exceed plain KNN, with class-wise OBL singled out as the strongest scheme. The motivation is that reflected points add geometry-aware, class-aligned coverage of sparse or high-dimensional feature spaces, at negligible computational cost.","feed_headline":"Reflecting data points lifts KNN on 26 datasets","feed_subtitle":"Simple opposite-sample augmentation sharpens decision boundaries in sparse, high-dimensional data.","key_machinery":"The load-bearing object is the opposition transformation $x^*_k = a_k + b_k - x_k$, a deterministic mirror reflection of a feature value about the midpoint of its range. Varying the source of the bounds ($a_k,b_k$)—global, class-specific, or local neighborhood—yields three augmentation schemes. Because the reflection is geometric rather than stochastic, the added points are claimed to expand the feature-space footprint of each class without injecting random noise.","core_discovery":"The central claim is that reflecting each training point through feature-wise boundaries, $x^*_k = a_k + b_k - x_k$, produces synthetic opposite samples that improve KNN classification when appended to the training set. The bounds $a_k,b_k$ are taken globally (Global OBL), per class (Class-Wise OBL), or from the $P$ nearest neighbors of the same class (Localized Class-Wise OBL). On 26 datasets, the paper reports that the OBL-augmented KNN variants beat both plain KNN and distance-weighted WKNN on accuracy and F1-score, and that Class-Wise OBL performs best. The authors attribute the gain to enriched decision boundaries and better coverage of underexplored regions, especially in high-dimensional and sparse settings.","pith_inferences":["The class-wise variant effectively mirrors each class separately, which should increase the density of minority classes relative to majority classes; if so, its advantage should be largest on imbalanced datasets, a pattern that could be tested by stratifying the 26 datasets by imbalance ratio.","A natural extension is to treat the bounds as learnable or to apply OBL repeatedly after different preprocessing stages, which might further improve separation.","The localized variant's neighborhood size $P$ is a new hyperparameter; its optimal value may scale with dimensionality, so an adaptive $P$ could be a testable improvement.","Combining OBL reflection with interpolation-based augmentation could produce a two-stage scheme that both expands the convex hull and adds opposite-region coverage; this is an editorial suggestion, not a paper claim."],"forward_implications":["If the reported gains hold under clean evaluation, OBL augmentation is a parameter-light way to improve KNN on high-dimensional and sparse datasets, with runtime overhead close to that of plain KNN.","Class-Wise OBL being the best of the three schemes suggests that class-specific bounds preserve or sharpen between-class structure better than global or local bounds.","Weighted KNN variants (WOBLKNN) add small additional gains in the reported tables, indicating that distance weighting and OBL augmentation are complementary.","The paper lists SVM, logistic regression, and gradient boosting as intended beneficiaries, so testing OBL augmentation on those classifiers is a direct next step if the KNN results are confirmed.","Because the opposite set is generated deterministically from the training data, it provides a reproducible augmentation that does not depend on random seeds or generative models."],"supporting_citations":[{"why":"Introduces Opposition-Based Learning, the reflection principle the paper adapts into a data transformation.","marker":"[7]"},{"why":"Cited as one of the OBL application references, supporting the claim that OBL improves optimization and learning across domains.","marker":"[8]"},{"why":"Cited alongside [8] for OBL's breadth of applications, grounding the paper's assertion that OBL accelerates convergence and improves solution quality.","marker":"[9]"},{"why":"Defines K-nearest neighbors as the proximity-based classifier the OBL augmentation is designed to boost.","marker":"[12]"},{"why":"Defines the distance-weighted KNN variant (WKNN) used as a baseline and as a weighted counterpart.","marker":"[16]"},{"why":"Source of the UCI machine-learning datasets used in the 26-dataset evaluation.","marker":"[49]"},{"why":"Source of the scikit-feature high-dimensional datasets used in the evaluation.","marker":"[50]"}],"fun_headline_variants":["Opposition-based reflection sharpens KNN on 26 datasets","Class-wise opposite samples boost KNN on 26 sets","Reflect data to refine KNN decision boundaries","Synthetic opposites improve KNN classification"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported gains assume all preprocessing statistics—Z-score parameters, chosen features, and OBL bounds—are computed from the training folds only, with the test fold never influencing them; the paper does not explicitly state that the train/test split occurs before these statistics are calculated.","fun_headline_variants_meta":{"raw":{"variants":["Opposition-based reflection sharpens KNN on 26 datasets","Class-wise opposite samples boost KNN on 26 sets","Reflect data to refine KNN decision boundaries","Synthetic opposites improve KNN classification"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000442,"raw_usage":{"total_tokens":2175,"prompt_tokens":816,"completion_tokens":1359,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":432,"completion_tokens_details":{"reasoning_tokens":1297}},"tokens_in":432,"tokens_out":1359,"duration_ms":12044,"temperature":1.0,"reasoning_tokens":1297,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:07:38.590087+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same 26-dataset protocol with a strict fold-wise pipeline in which Z-score normalization, mutual-information feature selection, and OBL min/max bounds are recomputed inside each training fold only, then compare OBL-KNN against plain KNN; if the accuracy advantage disappears or shrinks to statistical noise, the original gains came from leakage of test-set information into preprocessing.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Cited as one of the OBL application references, supporting the claim that OBL improves optimization and learning across domains."},{"cited_title":"R., & Salama, M","cited_arxiv_id":null,"evidence_quote":"Cited alongside [8] for OBL's breadth of applications, grounding the paper's assertion that OBL accelerates convergence and improves solution quality."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the distance-weighted KNN variant (WKNN) used as a baseline and as a weighted counterpart."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Source of the UCI machine-learning datasets used in the 26-dataset evaluation."},{"cited_title":"Differential evolution - based neural network training incorporating a centroid-based strategy and dynamic opposition-based learning","cited_arxiv_id":null,"evidence_quote":"Source of the scikit-feature high-dimensional datasets used in the evaluation."}],"review_version":1}