Sector-Mean: Deterministic Initialization of K-Means Centroids via Angular Sector Partitioning
Abhiyan Dhakal, Pranish Kafle, Rajani Chulyadyo
cs.LG
Quick verdict
Claims deterministic O(N) 2D K-Means seeding matching K-Means++ quality at 74.9% less init time; plausible, but Table IV silhouette stats look inconsistent.
Summary
The paper proposes Sector-Mean, a deterministic initialization scheme for K-Means restricted to two-dimensional data. Around the global centroid, the plane is split into k equal angular sectors and each sector's mean point becomes an initial centroid, giving O(N) single-pass seeding with a global-mean fallback for empty sectors. Across 15 two-dimensional datasets (SIPU A1–A3 and S1–S4, Birch1–3 at 100,000 points, UJIIndoorLoc and California Housing coordinates, plus Iris, Wine, and MNIST reduced to 2D by PCA), the authors compare Random, K-Means++, and Max-Min seeding under identical hand-implemented Lloyd iterations. Initialization time drops 74.9% versus K-Means++ and 59.8% versus Max-Min on average, and average iteration counts are the lowest of the four methods (~5% fewer than K-Means++, ~16% fewer than Max-Min). Friedman tests with Nemenyi post-hoc comparisons across the 15 datasets indicate statistically equivalent inertia and silhouette to K-Means++ and Max-Min, with significantly better ranks than Random. Main limitations: the method is defined only for 2D; deterministic competitors named in the related-work section (Var-Part, PCA-Part, angular predecessors) are never run; repetitions, seeds, and convergence tolerance go unreported; the appendix with ARI/NMI results is absent; and the silhouette row of Table IV contradicts the stated critical difference of 1.211.
Review complete
Assessment at a glance
6
risk flags identified
● high 1
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Sector-Mean deterministically initializes K-Means centroids in O(N) for 2D data by averaging points within k equal angular sectors around the global centroid.
- Average initialization time is reduced by 74.9% versus K-Means++ and 59.8% versus Max-Min across 15 datasets (mean of per-dataset reductions, Table II).
- Sector-Mean achieves the lowest average iteration count (~5% fewer than K-Means++, ~16% fewer than Max-Min) while quality metrics remain statistically tied with both baselines (Table IV).
- Sector-Mean differs significantly from Random initialization on inertia, iterations, and silhouette, which the paper reads as outperforming Random.
- The efficiency gain comes without compromising clustering quality, making Sector-Mean a practical deterministic alternative for 2D spatial clustering.
Strengths
Fully specified algorithm with explicit edge-case handling
Algorithm 1 plus Eqs. 1–5 specify global mean, arctan2 mapping, θ=π clamping, and the empty-sector global-mean fallback, making the method trivially reimplementable from the text alone
Unified protocol spanning three orders of magnitude in scale
Same hand-implemented stack for all methods, identical Lloyd iterations, 15 public datasets from 150 (Iris) to 100,000 points (Birch) and k from 3 to 100 (Sec. IV-A, Table I)
Correct non-parametric statistical framework
Honest scoping and counterexamples disclosed
Large-scale timing evidence consistent with complexity claims
Weaknesses
Method restricted to 2D, with high-dimensional benchmarks only tested after PCA projection
Closest related methods never compared empirically
Reproducibility gaps in the experimental protocol
Table IV silhouette row contradicts its own critical difference
No code release and missing appendix
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
Sector-Mean (arXiv:2609.06468v1) proposes a deterministic centroid initializer for K-Means on strictly 2D data: compute the global mean (Eq. 1), map every point to an angle via arctan2 (Eq. 2), slice the circle into k equal sectors (Eqs. 3–4, with θ=π clamped), and return per-sector means as seeds (Eq. 5; global-mean fallback for empty sectors). The authors claim (i) O(N) single-pass seeding, (ii) 74.9%/59.8% average initialization-time reductions vs K-Means++ and Max-Min, (iii) the lowest average iteration counts (~5%/~16% fewer), and (iv) statistically equivalent quality to K-Means++/Max-Min with significantly better ranks than Random. I verified the headline percentages arithmetically: 74.9% is the mean of per-dataset reductions from Table II, and the iteration means from Table III are 36.4 (SM) vs 38.0 (K-Means++) vs 43.1 (Max-Min) — the abstract's numbers hold.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is Sector-Mean initialization for K-Means?
A deterministic way to choose K-Means starting centers for two-dimensional data: compute the dataset's global center, convert every point's position to an angle around it, split the full circle into k equal angular sectors, and use the mean point of each sector as a starting centroid. It runs in a single O(N) pass and always returns the same seeds, unlike stochastic schemes such as K-Means++ or random selection.
Is Sector-Mean better than K-Means++?
According to the paper, it reaches the same clustering quality (inertia and silhouette are statistical ties with K-Means++ across 15 datasets) while initializing about 74.9% faster on average and converging in roughly 5% fewer iterations. The key trade-off: Sector-Mean is defined only for 2D data, whereas K-Means++ works in any dimension.
Does Sector-Mean work for high-dimensional data like images or text embeddings?
No. The method relies on arctan2 angles around a center point and is defined only for 2D points. The paper evaluates MNIST, Iris, and Wine only after projecting them to two dimensions with PCA and offers no extension to higher-dimensional spaces.
Is there code available for Sector-Mean?
The paper does not state any code or data release. However, the algorithm is short (13-line pseudocode) and simple to reimplement in NumPy, and all datasets used (SIPU, Birch, UJIIndoorLoc, California Housing, Iris, Wine, MNIST) are publicly available, so independent replication is feasible.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The appendix holding per-dataset ARI/NMI results and the critical-difference diagram (explicitly referenced in Secs. V-A4, V-B and VI-C) is not present in the provided text, so those external-validation results could not be checked.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.