{"id":"73e04e87-d9b6-4129-9e38-a8c9272175e2","arxiv_id":"2501.04099","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"NDESO moves noisy minority points toward their class centroids before random oversampling and is claimed to beat 14 resamplers, but its cross-validation protocol casts doubt on the headline result.","lead":"This paper proposes NDESO, a resampling method that shifts minority-class points whose nearest neighbors mostly belong to other classes toward their class centroid, then applies random oversampling. It claims higher average G-mean than 14 existing resamplers on 20 multiclass datasets, but the reported evaluation resamples before the train/test split, which can leak test information.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline comparison is invalid because Section 5.2 resamples each full dataset before the 80/20 split, letting test-set geometry and synthetic copies leak into training; the reported G-mean ranks therefore do not measure generalization.","rationale":"The reader's weakest assumption correctly identifies the most load-bearing flaw. The paper's own Section 5.2 explicitly describes resampling before partitioning, and every reported G-mean and mean rank is computed under that protocol. Since NDESO's displacement step uses global pairwise distances and labels, and since random oversampling duplicates points across the full dataset, the evaluation leaks test information into training. This is not a minor implementation detail; it is the foundation of the central claim that NDESO outperforms 14 baselines. A corrected inside-fold evaluation is the natural check, and it is feasible because the authors provide code and data. I also note the reader's secondary observations: the Nemenyi result puts NDESO in a non-significant block with Borderline-SMOTE, SMOTE-CDNN, and SMOTE-ENN, and the paper does not report fold-level variance. These reinforce the rejection, but the pre-split leakage alone is sufficient to invalidate the empirical claim as stated. The displacement rule in Eq. (5) also deserves scrutiny: it moves a point to distance phi_i from the centroid, which only moves it closer to the centroid when phi_i is smaller than the original centroid distance; otherwise it jumps past the point. However, that is a design issue, not the primary reason the comparison cannot be accepted.","tokens_in":27002,"tokens_out":2529,"duration_ms":28095,"concrete_test":"Re-run the released GitHub implementation with a corrected protocol: for each of the 20 datasets, perform a stratified train/test split first, fit each resampler only on the training fold, oversample within that fold, train each of the nine classifiers, and evaluate on the untouched test fold. Aggregate G-mean and recompute Friedman/Nemenyi ranks. If NDESO's average G-mean or mean rank falls materially below the reported 0.9362 and 1.85, or if its margin over Borderline-SMOTE, SMOTE-ENN, and SMOTE-CDNN is no longer significant, the central empirical claim is unsupported.","verdict_should_be":"REJECT","load_bearing_attack":"Section 5.2 states: \"A given dataset is tested using a resampling method to balance the class distributions. After resampling, the dataset is then partitioned into 80% training and 20% testing subsets using cross-validation.\" This is exactly the pre-split resampling protocol that standard imbalanced-learning evaluation avoids. In NDESO, the displacement step in Algorithm 1 computes pairwise distances over the entire dataset and relocates points judged noisy using labels from the full dataset; random oversampling then duplicates minority samples before the split. As a result, test instances can be repositioned before the test split, and synthetic copies of test instances can appear in the training set while the original instances remain in the test set. Classifiers such as k-NN, SVC, and tree ensembles then see near-duplicates of test points during training, inflating G-mean values and corrupting every rank used in the Friedman and Nemenyi tests. Because the paper's central claim is that NDESO achieves the highest average G-mean and the lowest mean rank across 20 datasets and nine classifiers, this protocol error is load-bearing: if the resampler is instead fitted only inside each training fold, the reported advantage may shrink or disappear. The same flaw also undermines the claimed statistical significance, since the Nemenyi test is applied to ranks derived from leaked evaluations.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes NDESO (Neighbor Displacement-based Enhanced Synthetic Oversampling), a hybrid resampler that first identifies points whose k-nearest neighbors are mostly from other classes, displaces those points toward their own class centroid, and then applies random oversampling to balance the class distribution. The authors evaluate NDESO against 14 baseline resamplers on 20 multiclass datasets with 9 classifiers, using G-mean and the Friedman/Nemenyi tests, and claim that NDESO achieves the highest average G-mean and the lowest mean rank. The core contribution is a displacement-based alternative to cleaning methods that delete noisy minority instances.","tokens_in":27166,"tokens_out":6205,"duration_ms":60018,"significance":"The idea of repositioning noisy minority points rather than deleting them is a plausible and potentially useful contribution to imbalanced multiclass resampling. The paper also provides a clear algorithmic description, a GitHub repository for reproducibility, and a broad comparison across datasets and classifiers. However, the main empirical claim is not supported by the experimental design as reported, because the evaluation protocol leaks test information into training. If the experiments were correctly re-run with the resampler fitted only on training folds, the rankings could change substantially. The current manuscript therefore does not establish the claimed superiority of NDESO.","major_comments":[{"comment":"The testing procedure resamples the full dataset before splitting: \"A given dataset is tested using a resampling method to balance the class distributions. After resampling, the dataset is then partitioned into 80% training and 20% testing subsets using cross-validation.\" In NDESO, Algorithm 1 computes pairwise distances over the entire dataset, determines displaceable points using labels from the full dataset, and then random oversampling duplicates minority instances before the split. This means test instances can be repositioned or their synthetic copies can be placed into the training set, while the original test instances remain in the test set. For classifiers such as k-NN, SVC, and tree ensembles, this inflates the reported G-mean values and invalidates all ranks used in the Friedman and Nemenyi tests. The resampler must be fitted only within each training fold. This is a load-bearing flaw: the abstract's claim that NDESO 'outperforms its competitors regarding average G-mean score and achieves the lowest statistical mean rank' is not supported by the current protocol.","section":"Section 5.2"},{"comment":"The Nemenyi test results as described contradict the claim of statistical superiority. The paper states that 'no significant differences exist among ... Borderline-SMOTE, SMOTE-CDNN, SMOTE-ENN, and NDESO' and then immediately asserts that 'with its lowest average mean rank of 1.85 ... it signifies better performance.' Being in the same non-significant block as three other methods means the observed rank advantage over those methods is not statistically significant. The wording must be corrected to report the actual inference: NDESO is statistically tied with those three methods and significantly better than the remaining ones. As written, the discussion overstates the statistical evidence.","section":"Section 5.3.2"},{"comment":"Many baseline methods have missing entries due to resampling failures (Table 2 lists 9 datasets for KMeans-SMOTE, 6 for SMOTE-ENN/ENN/ECDNN, etc.), yet average G-mean values and mean ranks are reported for all methods in Tables 3 and 5. The paper does not state how averages and ranks are computed in the presence of these missing values. If a method is ranked only on the datasets where it succeeded, different methods are compared on different subsets of datasets, which biases the comparison in favor of methods that always succeed (such as NDESO). The Friedman test also requires complete block data, so the reported p-value of 6.08e-20 is not interpretable unless the analysis was confined to the subset of datasets where all methods ran successfully or a valid incomplete-data method was used. The authors should either restrict all comparisons to the common subset of complete datasets, or use a paired analysis that explicitly accounts for missing values.","section":"Tables 3 and 5"}],"minor_comments":[{"comment":"The displacement formula x'_i = r_ci - S_v * phi_i moves the point beyond the centroid when the average neighbor distance phi_i exceeds the distance from x_i to the centroid. No condition is given to prevent the point from being relocated to the opposite side of the class center. The paper should specify the intended behavior or add clipping.","section":"Section 4.1, Eq. (5)"},{"comment":"The pseudocode sets R = X[y=cls] for each class and then refers to R[i] as the centroid for point i. Since R is a list of class-specific arrays, this indexing is not well-defined; the centroid for a point should be r_{c_i} as in Eq. (2). Please correct the notation.","section":"Algorithm 1, lines 14-17"},{"comment":"The number of cross-validation folds is computed as nsplits = min(5, min(ytrain)), but ytrain is defined only after the resampling step. Since resampling balances the classes, min(ytrain) will typically be the size of the majority class and will not reflect the original minority class size. Clarify the intended definition and the actual CV procedure used.","section":"Section 5.2"},{"comment":"Minor typographical issues include 'eucledian' in Algorithm 1 and the inconsistent formatting of class counts in Table 1 (e.g., the segment row has an unclosed bracket). These are cosmetic but should be cleaned up.","section":"General presentation"}],"recommendation":"reject","confidential_remarks":"The paper has a genuine idea, but the main empirical claim is invalidated by the pre-split resampling protocol, and the statistical interpretation contains an internal contradiction. A proper re-run of the experiments would be a substantial new study. I would not block the authors from resubmitting a corrected version, but the current manuscript does not meet the bar for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the NDE displacement idea is genuinely new and the paper is clearly written, but the evaluation protocol is broken. Section 5.2 resamples each full dataset before the 80/20 split, so test instances (and synthetic copies of them) can end up in the training set. That means the reported G-mean ranks are not estimates of out-of-sample performance, and the Friedman/Nemenyi conclusions built on those ranks don't hold.\n\nWhat's new: moving points whose k-neighborhood is majority other-class toward their class centroid before oversampling is a real mechanism, distinct from CDNN's remove/relabel. The method description is clear and the algorithm is simple to reproduce. The breadth of the benchmark (20 datasets, 9 classifiers, 14 baselines) is credible effort, and the failure table for baselines is honest.\n\nThe soft spots, in order of severity. First, the pre-split resampling leak. If the resampler is fitted inside each training fold—the standard protocol—the reported advantage may shrink or vanish. This is the load-bearing issue. Second, the Nemenyi interpretation is overstated: the test places NDESO in the same non-significant block as Borderline-SMOTE, SMOTE-CDNN, and SMOTE-ENN, yet the prose claims a clear significant margin. Those are incompatible. Third, Eq. (5) has an unstated overshoot condition: x' = r_c - S_v * phi_i moves the point closer to the centroid only if phi_i < distance(x_i, r_c); otherwise it jumps past the centroid. This deserves a sentence in the paper and possibly a clamping rule. Minor: only average G-mean is reported, no fold-level variance, and missing values in the tables (failed baselines) are not handled transparently in the ranking.\n\nWho this is for: anyone working on multiclass oversampling, especially SMOTE-family variants. The paper deserves a serious referee because the core idea is worth testing with a corrected protocol; the current version should be revised or re-run before the comparison can be accepted. I'd send it out, expecting heavy revision.","headline":"Interesting displacement-before-oversampling idea, but the evaluation resamples before the split, so the headline results don't measure generalization.","tokens_in":27789,"tokens_out":3436,"would_cite":false,"duration_ms":29231,"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 relocating noisy minority-class points toward their class centroids before random oversampling produces a multiclass resampler that outperforms 14 baselines on average G-mean across 20 datasets and nine classifiers.","keywords":["multiclass imbalance","oversampling","random oversampling","neighbor displacement","k-nearest neighbors","G-mean","Friedman-Nemenyi test","imbalanced classification"],"falsifier":"Re-run the same comparison with the resampling pipeline embedded inside each cross-validation fold, fitting the NDE displacement and random oversampling only on the training portion, and compare the resulting G-mean ranks. NDESO's claimed edge would be falsified if its mean rank no longer leads, or if its average G-mean advantage over SMOTE-ENN and SMOTE-CDNN shrinks below the Nemenyi critical difference.","tokens_in":26647,"feed_emoji":"🎯","tokens_out":6814,"duration_ms":62964,"temperature":0.7,"pith_summary":"The paper proposes NDESO, a hybrid resampling method for multiclass imbalanced classification. Its idea is to clean before balancing: find points whose k nearest neighbors are mostly from other classes, move them closer to their own class centroid by one average-neighbor-distance, and only then apply random oversampling. This preserves every point's class label while reducing overlap, which the authors argue avoids the information loss of undersampling and the noise amplification of SMOTE-style interpolation in sparse data. On 20 real-world datasets across nine classifiers, they report that NDESO achieves the highest average G-mean and the lowest mean rank among 15 resamplers, with the largest gains on the most extreme-imbalance datasets. If the comparison holds up, this is a low-cost, label-preserving preprocessing step that could be dropped into existing imbalanced-learning pipelines.","feed_headline":"Shift noisy points toward their class centers, then resample","feed_subtitle":"NDESO moves minority points closer to their centroids before random oversampling, beating 14 baselines on average G-mean.","key_machinery":"The load-bearing mechanism is the NDE displacement rule. For a point $x_i$ whose $k$-nearest-neighbor neighborhood is majority other-class, with class centroid $r_{c_i}$, normalized direction $\\vec{S}_v = (r_{c_i}-x_i)/S_i$, and $\\phi_i$ the average distance from $x_i$ to its $k$ neighbors, the point is moved to $x'_i = r_{c_i} - \\vec{S}_v\\,\\phi_i$. The rule compresses overlapping regions by pulling disputed points toward their own class center while keeping the original labels, and it requires only pairwise distances and a centroid, which is why it can run on very sparse classes. Random oversampling then balances the cleaned distribution.","core_discovery":"The central claim is that noisy minority-class points should be repositioned rather than removed or relabeled. A point counts as noisy when more than half of its k nearest neighbors belong to different classes; NDE then computes the average distance from that point to those neighbors and shifts the point along the unit vector toward its own class centroid by exactly that distance. Random oversampling then duplicates minority points until class sizes match the majority. In the authors' experiments this two-stage procedure outperforms 14 baseline resamplers, including SMOTE variants, ADASYN, and SMOTE-CDNN, in average G-mean and in Friedman/Nemenyi mean rank, and it runs successfully on sparse datasets where several baselines fail because a minority class has too few members.","pith_inferences":["The main risk to the ranking is evaluation leakage: the protocol resamples before splitting, so a leak-free re-run with the resampler fitted inside each fold is the decisive test of whether NDESO's G-mean advantage is real.","NDE could be treated as a generic denoising front end: composing it with oversamplers other than ROS, such as SMOTE or ADASYN, might inherit the cleaning benefit while adding synthetic diversity; the paper's own NDE+SMOTE experiments suggest the displacement, not the final sampler, drives much of the gain.","Because the displacement rule only needs a distance metric, the same cleaning idea could transfer to image, spatial, or graph data once a pairwise distance is defined; the paper names big-data and non-tabular resampling as open directions."],"forward_implications":["Datasets with an extreme imbalance ratio, up to 853:1 in the paper, are where NDESO shows its largest gains, suggesting the cleaning step matters most when minority classes are tiny.","Because NDESO displaces rather than deletes or relabels points, no minority class is erased by preprocessing, in contrast to undersampling or edited-neighbor methods that can collapse a class to zero members.","NDESO runs on sparse datasets where SMOTE-family samplers fail with neighbor-count errors, which broadens the set of multiclass problems a single resampler can handle.","The method's extra cost over plain random oversampling is the pairwise-distance computation of NDE, so the reported accuracy gain is obtained without a generative model."],"supporting_citations":[{"why":"Defines the SMOTE interpolation baseline whose overgeneralization and noise in sparse multiclass data motivate the displacement step.","marker":"[27]"},{"why":"SMOTE-CDNN, the closest hybrid baseline, which removes or relabels noisy points before oversampling; NDESO is positioned against it and outperforms it in the reported G-mean.","marker":"[28]"},{"why":"ADASYN, an adaptive oversampling baseline that NDESO is compared against and that fails on several sparse datasets.","marker":"[30]"},{"why":"Introduces centroid displacement-based k-nearest neighbors, the technique NDESO repurposes: instead of predicting labels by displacement, NDE uses displacement to clean points.","marker":"[32]"},{"why":"Supplies the Friedman and Nemenyi statistical testing framework used to convert G-mean results into the mean-rank claim.","marker":"[38]"},{"why":"Source for the G-mean metric definition used as the paper's primary performance measure.","marker":"[33]"}],"fun_headline_variants":["Move noisy minority points toward centroids, then oversample","Relocate noisy points closer to class centers before resampling","Shift noisy minority data to centroids, then random oversample","NDESO: move noisy points to centroid before oversampling","Neighbor displacement before oversampling boosts G-mean"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The experiments assume that resampling the whole dataset before the training/test split gives an unbiased estimate of generalization; if synthetic copies of test points can leak into training, the reported G-mean advantage may not reflect performance on new data.","fun_headline_variants_meta":{"raw":{"variants":["Move noisy minority points toward centroids, then oversample","Relocate noisy points closer to class centers before resampling","Shift noisy minority data to centroids, then random oversample","NDESO: move noisy points to centroid before oversampling","Neighbor displacement before oversampling boosts G-mean"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000492,"raw_usage":{"total_tokens":2365,"prompt_tokens":841,"completion_tokens":1524,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":457,"completion_tokens_details":{"reasoning_tokens":1452}},"tokens_in":457,"tokens_out":1524,"duration_ms":10679,"temperature":1.0,"reasoning_tokens":1452,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:41:18.210818+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the same comparison with the resampling pipeline embedded inside each cross-validation fold, fitting the NDE displacement and random oversampling only on the training portion, and compare the resulting G-mean ranks. NDESO's claimed edge would be falsified if its mean rank no longer leads, or if its average G-mean advantage over SMOTE-ENN and SMOTE-CDNN shrinks below the Nemenyi critical difference.","supporting_citations":[{"cited_title":"Wang, Stefanka S","cited_arxiv_id":null,"evidence_quote":"Introduces centroid displacement-based k-nearest neighbors, the technique NDESO repurposes: instead of predicting labels by displacement, NDE uses displacement to clean points."}],"review_version":1}