{"id":"6cbdf4d7-fb6f-42a5-ac2c-49982c76c648","arxiv_id":"2506.02366","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"GBABS samples only approximate borderline points detected via non-overlapping granular balls, reporting better classifier accuracy and noise robustness than GB-based and standard sampling baselines.","lead":"This paper introduces a data sampling method that builds small pure \"granular balls\" and keeps only the samples near class boundaries. The authors report that this compressed training set improves classifier accuracy and robustness to label noise in most of their tests.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The h=1 branch of RD-GBG's local-density center detection (Section IV-B1) deletes the nearest heterogeneous neighbor as class noise, which in clean data is typically a genuine boundary sample; the paper's own Fig.2 shows this happening in a noise-free dataset.","rationale":"The reader's weakest assumption identifies the local-density detection rules; I agree that this is the load-bearing component, but I sharpen it to a specific, checkable failure: the h=1 branch removes genuine boundary samples in the noise-free regime. This is more central than the complexity or statistical-reporting issues because it strikes at the method's stated purpose of preserving boundaries while removing noise. The paper's Fig.2, which presents a clean 4-class example, already shows a genuine boundary point being deleted as 'class noise' for (c3,l3). If that deletion occurs in real clean datasets, then the method is not doing approximate boundary sampling so much as aggressive boundary cleaning, and the claimed advantage over GGBS cannot be attributed to the proposed mechanism. The concern is testable by a simple synthetic experiment and an ablation of the h=1 rule. If the h=1 branch proves harmless or beneficial, the paper stands; if not, the central claim needs significant revision. I do not see this as a reason to reject outright: the empirical evidence and code are substantial, and the authors explicitly expose the rule, so a conditional acceptance with a demand for this ablation is appropriate. Other issues (unbounded iteration count t in the linear-time claim, missing standard deviations, mixed LightGBM results in Table IV) are secondary and can be addressed editorially. Verdict remains CONDITIONAL; my read does not change the reader's verdict, only adds a sharper condition.","tokens_in":21450,"tokens_out":11798,"duration_ms":119930,"concrete_test":"On a clean synthetic two-class dataset with known labels and no injected noise (e.g., two parallel Gaussian stripes with a small separating gap), run RD-GBG with rho=5 and record all samples removed by the h=1 and h=rho branches in Algorithm 1. Compute the fraction of removed samples that lie within the known boundary margin or are otherwise true (non-mislabeled) samples. Then ablate the h=1 branch: keep all other rules but never delete the nearest heterogeneous neighbor, and compare GBABS-DT accuracy on clean datasets. If h=1 deletes genuine boundary samples and removing it does not hurt (or improves) clean-data accuracy, the noise-detection rule is both destructive and unjustified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of GBABS—that it both samples borderline points and removes class noise without a purity threshold—rests on the local-density center detection rules in Section IV-B1. The rule states that when the nearest neighbor of a candidate center is heterogeneous, one computes h(c,l) via Eq. (2); if h=rho the center is deleted as noise, if h=1 the nearest heterogeneous neighbor is deleted as noise and the center becomes eligible, and if 1<h<rho the center is marked low-density. But a nearest heterogeneous neighbor in a clean dataset is exactly a genuine class-boundary point. Thus the h=1 branch deletes clean boundary structure under the label 'class noise.' The paper's own illustrative dataset (Fig. 2) is noise-free, yet for center (c3,l3), the nearest point (x1,y1) is heterogeneous while the other 4 of 5 neighbors are homogeneous, so (x1,y1) is removed as 'class noise.' This shows the detector cannot distinguish a true mislabel from a legitimate adjacent-class point. Because this deletion happens before any borderline sampling, the claimed boundary preservation is suspect: if clean boundary samples are removed, GBABS is not preserving the boundary it claims to approximate. Low-density points (1<h<rho) are not actually removed; they later become radius-0 orphan GBs (Algorithm 1 line 32), so the h=1 deletion is the main mechanism that discards boundary information. This is load-bearing for both the borderline-sampling contribution and the noise-robustness claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-stage sampling method for classification: RD-GBG, a restricted-diffusion granular-ball generation algorithm that avoids ball overlap and includes a local-density center detection rule for noise elimination, and GBABS, a granular-ball-based approximate borderline sampling method that selects samples near class boundaries. The authors claim that this is the first general sampling method that simultaneously performs borderline sampling and improves class-noise robustness, with linear time complexity and no need for an optimal purity threshold. Experiments on 13 datasets, with added class noise at 5%–40%, compare GBABS against GGBS, IGBS, SRS, SMOTE variants, and Tomek Links across five classifiers.","tokens_in":21847,"tokens_out":3793,"duration_ms":34304,"significance":"If the algorithm performs as claimed, it would be a practically useful contribution: it provides a general, compression-oriented sampling method that targets borderline points, and it comes with publicly available source code and a fairly broad experimental comparison. The RD-GBG overlap-avoidance mechanism is a reasonable idea, and the empirical evaluation spans multiple classifiers, datasets, and noise levels. However, the central noise-detection rule has a problematic behavior that undermines the boundary-preservation claim, and several experimental claims are stronger than the reported results support.","major_comments":[{"comment":"The local-density center detection rule deletes the nearest heterogeneous neighbor as class noise when h(c,l)=1. In a clean dataset, a nearest heterogeneous neighbor is typically a legitimate class-boundary point, not a mislabeled sample. The paper's own illustrative example in Fig. 2 is explicitly a noise-free dataset, yet for center (c3,l3) the nearest point (x1,y1) is removed as class noise. This behavior contradicts the paper's claim that the method preserves the class boundary while detecting noise, and it is load-bearing for both the borderline-sampling contribution and the noise-robustness claim. Please either revise the rule so that it can distinguish true label noise from genuine boundary structure, or provide direct evidence (e.g., on clean data with known boundaries) that this deletion does not degrade boundary fidelity.","section":"Section IV-B1, Eq. (2) and Fig. 2"},{"comment":"The claim that GBABS-based DT 'consistently achieves higher testingAccuracy' than plain DT is contradicted by Table II: plain DT has higher accuracy on S5 (0.8728 vs. 0.8709), S11 (0.9998 vs. 0.9994), and S12 (0.9750 vs. 0.9693). Similarly, Table IV shows GBABS-LightGBM at 5% and 10% noise (0.8660 and 0.8166) is worse than both GGBS-LightGBM (0.8690 and 0.8219) and plain LightGBM (0.8685 and 0.8222). The paper states only that GBABS 'generally performs better,' but the discussion in V-C and V-D conveys a stronger uniform superiority. Please analyze the datasets and settings where the method underperforms, and temper the claims accordingly.","section":"Section V-C, Table II and Section V-D, Table IV"},{"comment":"The time-complexity claim that the total cost is O(tqN + p m log m) and 'still linear' is not justified as stated. The number of iterations t and the number of classes q are not bounded by constants in Algorithm 1; in the worst case t can be as large as O(N) and q as large as O(N). The text merely says the total is 'much lower than O(tqN),' which is not a formal bound. Please provide a rigorous worst-case analysis or clearly state the assumptions under which the method is linear in practice, supported by measured runtimes.","section":"Section IV-B3 and Section IV-C"}],"minor_comments":[{"comment":"GGBS and IGBS are introduced in Section III-B as proposed in reference [23], but Section V-A1 cites [27] for these baselines; please correct the citation to avoid ambiguity.","section":"Section V-A1 and Section III-B"},{"comment":"There is a typo in 'Accuray' that should be 'Accuracy'.","section":"Section V-A3"},{"comment":"In the discussion of Fig. 6(a), the text says that for datasets with unclear class boundaries such as S3 the sampling ratio of GBABS is slightly higher than that of GGBS, but the figure appears to show GBABS with a lower ratio for S3; please verify the figure labels or the text.","section":"Section V-B"},{"comment":"The conclusion states that 'the time complexity of the GBABS is not ideal when facing high-dimensional feature spaces,' which appears to contradict the earlier claim that the total complexity is linear; please reconcile these statements.","section":"Section VI"},{"comment":"There is a typo in the notation: 'based on(c_j, l_i),(c_k, l_k)' should probably be 'based on(c_j, l_j),(c_k, l_k)'.","section":"Algorithm 2, line 5"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of the journal and the code release is a strength. However, the main technical concern is that the noise-detection rule in Section IV-B1 deletes genuine boundary samples in clean data, which directly affects the paper's central boundary-preservation claim. The experimental results also show several cases where the proposed method is not superior, especially for LightGBM and some standard datasets. These issues are addressable with a revised algorithm or with a more careful empirical analysis, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Xiang,\n\nThis is a useful, honest piece of work with one load-bearing conceptual problem. The new contribution is real: RD-GBG's overlap-free ball construction and the per-dimension heterogeneous-neighbor rule for identifying borderline centers are not in the earlier GBS/GBG++ line, and the code is public. The experiments are broad: 13 datasets, five classifiers, noise up to 40%, and the standard-data DT comparison includes a Wilcoxon test. The parameter sensitivity analysis for rho is also a plus.\n\nThe soft spot is the one the stress-test flagged, and it is not manufactured. Section IV-B1's h=1 rule deletes the nearest heterogeneous neighbor as class noise whenever exactly one of a candidate center's rho neighbors is heterogeneous. In a noise-free dataset, that neighbor is precisely a genuine class-boundary point. The paper's own Fig. 2 shows this happening: (x1,y1) is removed as \"class noise\" in a clean four-class dataset. So the noise detector cannot distinguish a true mislabel from a legitimate adjacent-class sample. Because this deletion happens before any borderline sampling, the claim that GBABS preserves class boundaries while removing noise is internally inconsistent. The method may still work as a heuristic—the accuracy results say it often does—but the \"outstanding on class noise\" story is not supported by the mechanism as described.\n\nOther issues are more ordinary. The rho value used in the main tables is never reported, which hurts reproducibility. The noise-experiment comparisons in Table IV have no error bars or significance tests; GBABS-LightGBM actually loses to GGBS-LightGBM at 5% and 10% noise, so \"generally outperforms\" is overstated. The linear time complexity claim is not rigorous because t, the number of iterations, is unbounded; worst case is O(tqN) with t possibly O(N). And the two main baselines, GGBS and IGBS, are from the same group—not disqualifying, but the independent baselines (SRS, SMOTE variants, Tomek) are doing a lot of the grounding work.\n\nBottom line: a genuine algorithmic contribution, openly implemented, with useful empirical coverage. It deserves a serious referee, but the referee should push hard on the boundary/noise conflation and require complete experimental reporting. If those are fixed, it is a reasonable addition to the sampling and granular-computing literature. I would send it out.","headline":"A genuine new GB-based borderline sampler with a real design flaw: its noise detector deletes clean boundary points, so the noise-robustness story needs rework, but the paper deserves peer review.","tokens_in":22300,"tokens_out":2861,"would_cite":false,"duration_ms":25695,"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":"GBABS is a granular-ball sampling method that targets borderline samples, removes class noise during ball generation, and claims linear-time compression with higher classifier accuracy than existing GB-based and representative baselines.","keywords":["granular-ball computing","borderline sampling","class noise","data sampling","undersampling","imbalanced classification","granular-ball generation","classification"],"falsifier":"Take a synthetic two-class dataset with known injected label flips, run RD-GBG with a fixed $\\rho$, and compare the samples flagged as class noise under the $h(c,l)=\\rho$ rule against the ground-truth flips. If precision on the flips is no better than random selection, or if most flagged samples are actually boundary points, the claimed noise robustness without a purity threshold collapses.","tokens_in":21197,"feed_emoji":"🎯","tokens_out":9531,"duration_ms":84473,"temperature":0.7,"pith_summary":"This paper tries to establish that borderline sampling and class-noise cleaning can be done together by one general, classifier-agnostic sampling method. It proposes RD-GBG, a granular-ball generation procedure that builds pure, non-overlapping balls by restricted diffusion, along with GBABS, which samples only the samples adjacent to class boundaries. If the claim holds, training-data compression, robustness to label noise, and imbalanced-class performance all improve without tuning a purity threshold. The reported experiments support the claim on 13 datasets, with GBABS-based classifiers outperforming GB-based sampling and representative baselines, especially at high noise ratios.","feed_headline":"Granular-ball sampler keeps boundaries, beats label noise","feed_subtitle":"GBABS drops class noise and redundant points, lifting accuracy on 13 datasets from 5 to 40 percent noise.","key_machinery":"The carrying mechanism is restricted diffusion-based granular-ball generation (RD-GBG) paired with dimension-wise borderline detection. A granular ball is a ball with center $c$, radius $r$, label $l$, and the sample set $O$ it covers; RD-GBG makes every ball pure by expanding from a local-density center until the $\\omega$-th nearest neighbor is homogeneous and the $(\\omega+1)$-th is heterogeneous, then caps the radius at the conflict radius $r_{\\mathrm{conf}}(c)$ if the ball would overlap an earlier one. This builds a non-overlapping cover of the data whose boundaries match the class structure, avoiding the blurred or shrunken boundaries that overlapping balls cause. For sampling, the centers of all balls are ordered along each feature; a center is borderline if a left or right neighbor in any dimension belongs to a different class, and the samples adjacent to those centers in the corresponding dimension are collected. Local-density center detection, which counts heterogeneous neighbors $h(c,l)$ among the $\\rho$ nearest neighbors, acts as the built-in noise filter: it labels candidate centers or their nearest neighbors as class noise and returns low-density samples to the undivided set.","core_discovery":"On its own terms, the paper claims that a granular-ball sampling method can do two things existing general samplers do not: keep the samples that carry class-boundary information and remove class noise from the training set, with no purity threshold to tune. The method first runs RD-GBG, which grows pure, non-overlapping granular balls from local-density centers by restricted diffusion: each ball expands until it meets a heterogeneous sample or a previously generated ball, and if the locally consistent radius $\\mathrm{CR}(c)$ would collide with an existing ball, the radius is truncated to the restricted maximum consistent radius $r_{\\max}(c)$. Then GBABS projects the ball centers onto each feature dimension and marks as borderline any ball whose left or right neighbor in that dimension has a different label; the samples just across those boundaries form the sampled set. On 13 datasets with injected class noise from 5% to 40%, the authors report that classifiers trained on GBABS samples beat GB-based GGBS, simple random sampling, and the unsampled data, and that the advantage grows as noise rises. The paper also reports lower sampling ratios than GGBS and stable behavior as the density tolerance $\\rho$ ranges from 3 to 19, which it reads as evidence that no optimal purity threshold is required.","pith_inferences":["A natural extension is to lift RD-GBG's local-density center detection out of sampling and use it as a standalone label-noise filter, since it requires only class-specific nearest-neighbor counts.","The boundary-focused selection could also serve as a preprocessing step for active learning or large-scale support-vector training, where only points near the decision surface matter.","At high noise ratios on small datasets, pure borderline undersampling may remove too many minority samples; pairing GBABS with a constrained oversampling step is a testable extension that its own G-mean rankings hint at.","One direct check of the method's geometric claim is to compare the samples GBABS selects with the support vectors of a support-vector machine trained on the same data; high overlap would confirm that the ball centers approximate the true decision boundary."],"forward_implications":["GBABS should keep classifiers accurate under class noise: the reported average accuracy at 40% noise is higher for GBABS than for GGBS, SRS, or the full dataset on every tested classifier.","GBABS should compress more aggressively than GGBS on almost all datasets, and the gap widens as noise increases, because noise removal and borderline-only sampling both cut samples.","On imbalanced data, GBABS should rank at or near the top in G-mean among eight sampling methods, particularly on noisy imbalanced datasets, because it undersamples the majority more aggressively near boundaries.","The density tolerance $\\rho$ should be an insensitive hyperparameter: sampling ratio and decision-tree accuracy stay roughly flat for $\\rho$ from 3 to 19.","The total runtime should stay roughly linear in the number of samples, $O(tqN + p m \\log m)$, so the method scales to large datasets."],"supporting_citations":[{"why":"Defines the granular-ball representation, purity measure, and splitting idea that RD-GBG reworks.","marker":"[22]"},{"why":"Introduces the GGBS and IGBS sampling methods that GBABS treats as its main GB-based baselines.","marker":"[23]"},{"why":"Provides the k-division granular-ball generation whose overlap and purity-threshold sensitivity RD-GBG aims to fix.","marker":"[27]"},{"why":"Supplies the synthetic oversampling baseline used in the imbalanced-data comparisons.","marker":"[39]"},{"why":"Supplies a borderline-focused oversampling baseline that GBABS contrasts with its own boundary sampling.","marker":"[12]"},{"why":"Supplies the nearest-neighbor undersampling baseline for the imbalanced experiments.","marker":"[16]"},{"why":"Supplies simple random sampling, the unbiased general sampling baseline matched to GBABS's sampling ratio.","marker":"[17]"}],"fun_headline_variants":["Granular-ball sampler boosts accuracy under label noise","GBABS: keep boundaries, drop noise, beat samplers","No purity threshold needed for noisy-data sampling with GBs","Approximate borderline sampling that filters class noise"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a fixed rule on nearest-neighbor counts can tell genuine class noise from useful borderline points: if all $\\rho$ nearest neighbors are foreign the center is noise, if exactly one is foreign that neighbor is noise, and otherwise the center is a low-density sample, with no theoretical or held-out justification for those cutoffs.","fun_headline_variants_meta":{"raw":{"variants":["Granular-ball sampler boosts accuracy under label noise","GBABS: keep boundaries, drop noise, beat samplers","No purity threshold needed for noisy-data sampling with GBs","Approximate borderline sampling that filters class noise"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00031,"raw_usage":{"total_tokens":1803,"prompt_tokens":1016,"completion_tokens":787,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":632,"completion_tokens_details":{"reasoning_tokens":723}},"tokens_in":632,"tokens_out":787,"duration_ms":6748,"temperature":1.0,"reasoning_tokens":723,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:25:27.260525+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a synthetic two-class dataset with known injected label flips, run RD-GBG with a fixed $\\rho$, and compare the samples flagged as class noise under the $h(c,l)=\\rho$ rule against the ground-truth flips. If precision on the flips is no better than random selection, or if most flagged samples are actually boundary points, the claimed noise robustness without a purity threshold collapses.","supporting_citations":[{"cited_title":"Granular ball sampling for noisy label classification or imbalanced classification,","cited_arxiv_id":null,"evidence_quote":"Introduces the GGBS and IGBS sampling methods that GBABS treats as its main GB-based baselines."},{"cited_title":"An efficient and adaptive granular-ball generation method in classification problem,","cited_arxiv_id":null,"evidence_quote":"Provides the k-division granular-ball generation whose overlap and purity-threshold sensitivity RD-GBG aims to fix."},{"cited_title":"Two modifications of cnn,","cited_arxiv_id":null,"evidence_quote":"Supplies the nearest-neighbor undersampling baseline for the imbalanced experiments."}],"review_version":1}