Pith. sign in

REVIEW 4 major objections 8 minor 9 references

Anomaly Detection and Improvement of Clusters using Enhanced K-Means Algorithm

T0 review · 4 major / 8 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read The paper claims an enhanced k-means algorithm that iteratively removes points far from cluster centroids and re-centers until intra-cluster variance converges, yielding tighter clusters and an explicit outlier set in one pass.

desk verdict A k-means outlier-trimming variant whose reported gains are probably artifacts of scoring only the surviving points; the evaluation never states what data the metrics are computed on. read the letter →

arxiv 2505.24365 v1 pith:3Q3D2GED submitted 2025-05-30 cs.LG cs.PF

classification cs.LGcs.PF
keywords MachineLearningClusteringUnsupervisedK-meansAnomalyDetectionOutlierRemovalIntra-clusterVarianceChebyshev'sInequality
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that a modified k-means procedure can do two jobs at once: detect anomalies and produce tighter, better-separated clusters. After a standard k-means pass, the algorithm measures each point's Euclidean distance to its cluster centroid, deletes every point more than two standard deviations from the cluster's mean distance, then recomputes centroids and repeats until average intra-cluster variance stops changing. The authors report that this iterative trimming lowers variance by 18.7% on synthetic data and 88.1% on the Wine Quality dataset, with accuracy and F1 rising 22.5% and 20.8% on that same real dataset. The practical interest is that clustering and outlier removal, usually handled separately, become one procedure with a single interpretable threshold built on density around centroids rather than distribution tails.

What carries the argument

The central object is the Enhanced K-Means loop, which combines an average intra-cluster variance objective $V_{\mathrm{avg}}$ with a Chebyshev-inequality threshold. For each cluster, it computes the array of Euclidean distances from centroid to each point, sets the cutoff at the mean distance plus two standard deviations, and deletes points beyond it, relying on Chebyshev's inequality with $k=2$ to guarantee that at least 75% of points are retained as non-outliers. Recomputing centroids and repeating until the change in variance $\Delta V = V_n - V_{n-1}$ approaches zero is what converts ordinary k-means into a simultaneous cluster refiner and anomaly detector.

What would settle it

Run the algorithm on a labeled dataset, then compute all reported metrics on the full original dataset with removed points assigned to their nearest surviving cluster; if the variance reduction and accuracy gains shrink or vanish, the claimed improvement is an artifact of trimming rather than a real clustering gain.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is an iterative refinement of k-means: each cluster's points are ranked by distance from their centroid, points beyond two standard deviations from the mean distance are labeled outliers and removed, centroids are recomputed on the survivors, and the loop repeats until the average intra-cluster variance converges to what the authors call a global minimum. The removed points form an explicit outlier set, so a single run yields both cleaned clusters and an anomaly list. Measured against standard k-means, the algorithm reports variance reductions of 18.7% on synthetic data, 57.9% on Breast Cancer Wisconsin, and 88.1% on Wine Quality, together with improved silhouette, Calinski-Harabasz, and Davies-Bouldin scores and improved label-agreement metrics.

Load-bearing premise

The central assumption is that every point more than two standard deviations from its cluster centroid's mean distance is a genuine outlier, and that removing it improves clusters even when the evaluation is done on the data that survived removal.

Editorial extensions

If this is right

  • A single run of this algorithm gives both cleaned clusters and an explicit anomaly set, so a data pipeline can skip a separate outlier-detection stage.
  • The reported Wine Quality result implies that standard k-means clusters on multi-feature, multi-label data can be tightened substantially by deleting roughly 7.6% of points.
  • The Chebyshev-derived threshold puts the cleaning budget under user control, since changing k changes the fraction of points retained.
  • The Breast Cancer results show the method maintains or improves supervised metrics while removing about 10.1% of points, suggesting noise removal need not sacrifice classification agreement.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the paper does not state whether accuracy and F1 are computed on the full original dataset or only on the points that survived outlier removal; if the latter, the gains are partly mechanical and a held-out evaluation is needed to confirm they generalize.
  • Editorial inference: the distance-only machinery should transfer to centroid-based methods besides k-means, such as k-medoids, because it never uses k-means-specific updates.
  • Editorial inference: a natural testable extension is an adaptive Chebyshev bound chosen per cluster or per dataset, which the paper itself lists as future work and which could be validated by cross-validated cluster-assignment quality.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 8 minor

Summary. The paper proposes an iterative enhancement of K-means that, after initial clustering, removes from each cluster the points whose distance to the centroid exceeds the cluster mean plus two standard deviations, then recomputes centroids and repeats until convergence. The authors claim this reaches a 'global minimum' of intra-cluster variance and report variance reductions of 18.7% (synthetic) and 88.1% (Wine Quality), along with accuracy and F1 gains of 22.5% and 20.8% on the Wine Quality dataset. Evaluation uses intrinsic measures (silhouette, Calinski-Harabasz, Davies-Bouldin) and extrinsic measures (Jaccard, V-measure, accuracy, F1) on synthetic, Wisconsin Breast Cancer, and Red Wine Quality datasets.

Significance. If the results were valid, the method would be a simple and portable way to obtain tighter clusters and to flag anomalies, and the Chebyshev-based threshold selection is a reasonable practical heuristic. However, the evaluation protocol is never specified: it is unclear whether metrics are computed on the full data or on the trimmed subset, and if the latter, the reported improvements are mathematical consequences of deleting the points farthest from centroids. The convergence claim is unproven, the extrinsic evaluation lacks essential details, and no comparison is made to the existing outlier-removal clustering algorithms cited in the paper. The manuscript therefore provides no validated evidence for its central claims.

major comments (4)
  1. [§VI, Figs. 6-7 and Algorithm 1] Figures 6 and 7 report large improvements in intra-cluster variance, silhouette, Calinski-Harabasz, Davies-Bouldin, accuracy, and F1, but the text never states whether these metrics are computed on the full original dataset or only on the points that survive the outlier-removal loop (Algorithm 1, lines 8-11). Because the algorithm deletes every point whose distance to its centroid exceeds the cluster mean plus two standard deviations, computing any intrinsic metric on the surviving points guarantees a decrease in intra-cluster variance and removes exactly the points most likely to be misclassified. The headline numbers (88.1% variance reduction, 22.5% accuracy gain, 20.8% F1 gain on Wine Quality) therefore do not demonstrate improved clustering of the original dataset unless the evaluation is performed on held-out or full data. The authors must state explicitly where the metrics were computed and redesign the experiments if they were computed on the trimmed subset.
  2. [§III, Eq. (7) and Algorithm 1] The abstract and Section III claim that the iterative process 'reduces the intra-cluster variance until the variance approaches a global minimum.' This claim is not supported: Eq. (7) is written as 'lim_{a→0} : a = ΔV[V_n, V_{n-1}]', which is not a well-formed mathematical statement, and Algorithm 1's while-loop condition is never defined. The procedure is a greedy, iterative trimming of extreme points; deleting points cannot increase the variance of the remaining points, but the sequence of centroids could oscillate or converge to a local optimum of the trimmed objective, not a global minimum. A precise stopping criterion and either a proof of convergence or an empirical analysis with multiple restarts are needed before this central claim can be accepted.
  3. [§V and §VI] The extrinsic evaluation is not reproducible. The paper reports accuracy, F1, Jaccard, and V-measure against ground-truth labels but does not explain how the cluster labels are matched to class labels (e.g., Hungarian assignment), how the Jaccard coefficient is computed for clustering (pairwise vs. set), or what value of β is used in Eq. (11). In addition, the paper cites existing outlier-removal clustering methods in Section II (ORC, KMOR, KMN, MORe++) but never compares the enhanced K-means against them. Reporting improvements over plain K-means is insufficient to establish the method as a contribution to the outlier-detection or robust-clustering literature.
  4. [§VI and §V] All real-data experiments use a fixed number of clusters (k=5) with no sensitivity analysis or justification, despite Figure 4 showing that the method's benefit decreases as k grows on synthetic data. The WBC and Wine Quality results are therefore single-k instances and do not support the general claim that the algorithm 'produces tighter clusters than standard k-means' across datasets. Additional k values, or an explanation of how k was selected, are required.
minor comments (8)
  1. [§III, Eq. (7)] The convergence expression is not a valid equation; define ΔV and state a concrete tolerance or stopping threshold.
  2. [§VI, Eq. (11)] The V-measure formula uses an unspecified β; the standard V-measure uses β=1, so clarify or define β.
  3. [§II] Several sentences are missing mathematical symbols, e.g., 'Depth-based outlier-detection algorithms work well for [8,9], but become inefficient for .' and 'This had a complexity of , where is the dimensionality and is the number of objects.'
  4. [§VI] The 'Boldmin index' should be the 'Davies-Bouldin index'.
  5. [§VII] The phrase 'V on Mises-Fisher distribution' should be 'von Mises-Fisher distribution'.
  6. [Algorithm 1] The pseudocode does not specify the stopping condition for the while loop or the convergence tolerance, and the referenced line numbers 8-11 for outlier removal are not visible in the pseudocode.
  7. [§V] The synthetic dataset is described only as 'a thousand randomly generated points'; the underlying distribution should be specified to make the results interpretable.
  8. [References] There are formatting inconsistencies, e.g., 'V . Hodge' and missing italics, and the bibliography should be cleaned up.

Circularity Check

2 steps flagged · score 8.0 of 10

Variance 'improvements' are the algorithm's own deletion objective; supervised gains may be scored post-trimming, with no held-out evaluation.

  1. fitted input called prediction [Section III (Proposed Algorithm), Equations (3)-(7), Algorithm 1; results in Sections V and VI]
    "all points greater than 2 standard deviations from the mean are labelled as outliers and removed from the respective cluster. ... Once the outliers are removed, the centroids are re-positioned to the modified dataset, and the outliers are then removed from that modified dataset using the same algorithm depicted above. This process is iterated until the variance converged to a global minimum."

    The paper reports variance reductions of 18.7% and 88.1% as evidence of tighter clusters. But the average intra-cluster variance defined in Eq. (4) is exactly the quantity minimized by Algorithm 1: the deletion rule removes the points with the largest distances from their centroids, which removes the largest terms in the variance sum of Eq. (3). Recomputation of centroids on the survivors only decreases the same objective further. Therefore the headline variance reductions are consequences of the outlier-removal operation, not independent measurements of cluster quality on the original dataset.

  2. fitted input called prediction [Section VI (Real Data), Figures 6B and 7B and accompanying text]
    "The algorithm found 121 outliers in the dataset, which is 7.6% of the dataset. ... The algorithm was able to increase the correct classifications made by the algorithm as the accuracy and F1 metric increased by 22.5% and 20.8% respectively."

    The paper never states whether the supervised metrics are computed on the full original dataset or only on the points that survive the outlier-removal loop. Algorithm 1 deletes points whose distance to their assigned centroid exceeds the cluster mean plus two standard deviations, i.e., the most peripheral and typically most misclassification-prone points. If Figures 6B and 7B are scored on the cleaned subset, the accuracy and F1 gains are at least partly forced by the trimming operation rather than by any learned improvement in cluster structure. No held-out evaluation is described, so the central empirical claim is not independently verified.

full rationale

The paper's central assertion is that the enhanced k-means algorithm yields tighter clusters and better classification. For the intrinsic claim, the reasoning is circular by construction: the algorithm explicitly minimizes average intra-cluster variance by discarding the highest-variance points and then reports the resulting variance decrease (18.7% on synthetic data, 88.1% on Wine Quality) as a positive result. That is not an independent evaluation; it is a restatement of the deletion rule in Algorithm 1 and the variance definition in Eqs. (3)-(4). For the extrinsic claim, the evaluation protocol is not specified: the paper reports accuracy and F1 improvements of 22.5% and 20.8% on the Wine Quality dataset without saying whether those scores are computed on the full original data or on the 92.4% of points that remain after removing 121 outliers. If the latter, the gains are an artifact of removing the points most likely to be misclassified. There is no self-citation load-bearing argument in the paper, and no imported uniqueness theorem; the circularity is empirical rather than citational. Because the headline variance result is forced by the algorithm's own objective, and the supervised result is at best unverified and at worst similarly forced by trimming, a score of 8 is appropriate.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central results depend on a hand-chosen trimming threshold, a manually set cluster count, and an unverified assumption that deleting far points cleans real noise rather than removing valid structure.

free parameters (3)
  • Outlier removal threshold = 2 standard deviations (≥75% non-outlier fraction via Chebyshev)
    The 75% floor is chosen by the authors, not derived from data; changing it changes the outlier count and all reported metrics.
  • Number of clusters k = 5
    Set manually for all real-data experiments; no selection criterion is given.
  • Convergence tolerance a = not specified
    The stopping condition in Eq. (7) uses a limit a→0, but the actual threshold used in experiments is not reported.
assumptions (4)
  • standard math Chebyshev's inequality applies to the distribution of point-centroid distances
    Used to justify the 2-sigma threshold in Section III, valid for any distribution with finite variance.
  • domain assumption The distance array is not Gaussian
    Stated without evidence in Section III; irrelevant to Chebyshev bound but used to motivate the choice.
  • ad hoc to paper Removing points far from centroids corresponds to removing anomalies
    No ground-truth anomaly labels are used; the paper assumes cluster-distance anomalies are real outliers.
  • domain assumption Class labels align with cluster structure
    Extrinsic evaluation maps k-means clusters to class labels via majority vote or similar, but the labeling procedure is not described.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Anomaly Detection and Improvement of Clusters using Enhanced K-Means Algorithm." pith.science (2026). https://pith.science/paper/3Q3D2GED

@misc{pith2026250524365,
  author       = {Pith},
  title        = {Pith review of: Anomaly Detection and Improvement of Clusters using Enhanced K-Means Algorithm},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3Q3D2GED}},
  note         = {Machine review of arXiv:2505.24365}
}
read the original abstract

This paper introduces a unified approach to cluster refinement and anomaly detection in datasets. We propose a novel algorithm that iteratively reduces the intra-cluster variance of N clusters until a global minimum is reached, yielding tighter clusters than the standard k-means algorithm. We evaluate the method using intrinsic measures for unsupervised learning, including the silhouette coefficient, Calinski-Harabasz index, and Davies-Bouldin index, and extend it to anomaly detection by identifying points whose assignment causes a significant variance increase. External validation on synthetic data and the UCI Breast Cancer and UCI Wine Quality datasets employs the Jaccard similarity score, V-measure, and F1 score. Results show variance reductions of 18.7% and 88.1% on the synthetic and Wine Quality datasets, respectively, along with accuracy and F1 score improvements of 22.5% and 20.8% on the Wine Quality dataset.

Figures

Figures reproduced from arXiv: 2505.24365 by the authors.

Figure 1
Figure 1. The image depicts how K-means clusters are affected by anomalies. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 3
Figure 3. The image shows the local and global outliers detected by the algorithm for 5 clusters [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 2
Figure 2. The image shows the evaluation of the model based on normalized [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figures from the paper (3 more)
Figure 5
Figure 5. Figure 5: Box plot showing the ratio of the distance of an outlier to the average [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Evaluation of algorithm on WBC dataset, with number of centroids set as 5. A shows the intrinsic measures. B shows the extrinsic measures. Figure A clearly shows the efficiency and effectiveness of the algorithm in unsupervised learning whilst using a real dataset. The…
Figure 7
Figure 7. Figure 7: Evaluation of algorithm on Red Wine Quality dataset, with number of centroids set as 5. A shows the intrinsic measures. B shows the extrinsic measures [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

9 extracted references · 9 canonical work pages

  1. [1]

    D. M. Hawkins, Identification of Outliers. Springer Netherlands, 1980

  2. [2]

    A Survey of Outlier Detection Methodologies,

    V . Hodge and J. Austin, “A Survey of Outlier Detection Methodologies,” Artificial Intelligence review, vol. 22, pp. 85–126, 2004

  3. [3]

    P. J. Rousseeuw and A. M. Leroy, Robust regression and outlier detection. Wiley-Interscience, 2003

  4. [4]

    Ester, M., Kriegel, H.-P., Sander, J., Xu, X., A density-based algorithm for discovering clusters in large spatial databases with noise, KDD, 1996

  5. [5]

    Outliers in statistical data

    Barnett V ., Lewis T.: “Outliers in statistical data”, John Wiley, 1994

  6. [6]

    Preparata F., Shamos M.: “Computational Geometry: An Introduction “, Springer, 1988

  7. [7]

    Exploratory Data Analysis

    Tukey J. W.: “Exploratory Data Analysis”, Addison-Wesley, 1977

  8. [8]

    Ruts I., Rousseeuw P.: “Computing Depth Contours of Bivariate Point Clouds, Journal of Computational Statistics and Data Analysis, 23, 1996, pp. 153-168

Show all 9 references
  1. [9]

    K-means clustering with outlier removal

    E. Knorr and R. Ng. Algorithms for mining distance-based outliers in large datasets. In VLDB, 1998. 10.Knorr, E., Ng, R. & Tucakov, V . Distance-based outliers: algorithms and applications. The VLDB Journal 8, 237–253 (2000). https://doi.org/ 10.1007/s007780050006 11.Gan, Guoj...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.