{"id":"5e55ad15-8b1f-4428-9db2-7d56ccac361f","arxiv_id":"2411.08867","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An elbow in how sorted GLOSH scores change across density settings selects a near-best minpts, and a regression-adjusted knee in the score list sets the inlier/outlier threshold, with no labels required.","lead":"This paper adds two automatic settings to the GLOSH outlier detector, so a user no longer has to choose a density parameter or a cutoff score by hand. The work matters for anomaly detection in settings where nothing is known about the data in advance, though the method is tested only on synthetic outliers.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ORD-Profile is computed on sorted score lists, so it cannot detect the ranking changes that P@n depends on; the claimed elbow-to-best-P@n link may be an artifact of score-distribution convergence.","rationale":"The reader's weakest assumption was that the elbow-to-best-P@n regularity is empirical rather than derived, and that it weakens for local outliers. My concern is more specific: the ORD-Profile, as defined, measures dissimilarity between sorted score distributions, not between rankings, so it is not even the right proxy for the quantity the central claim depends on. This is a load-bearing gap because P@n is a ranking quality measure and the method's only signal is a distributional dissimilarity. The concern is not an accusation of dishonesty; the authors are candid about the empirical nature of the observation and about local-outlier difficulties. The proposed concrete test would settle the issue by comparing the ORD-Profile elbow against both P@n and a ranking-aware dissimilarity on independent data. If the test fails, the paper's core contribution would need substantial revision; if it passes, the empirical regularity would be validated despite the weak theoretical justification. I therefore keep the reader's CONDITIONAL verdict unchanged: the paper is promising but not yet established, and the missing ranking-aware validation is the decisive condition.","tokens_in":20175,"tokens_out":6025,"duration_ms":66649,"concrete_test":"On each held-out dataset (e.g., ADBench synthetic benchmarks generated with a different recipe, not just Banana), compute for every minpts in [2,100]: (i) GLOSH P@n, (ii) the ORD-Profile value Delta(S_k,S_{k+1}) as defined in Eq. 4, and (iii) a ranking-aware dissimilarity between consecutive minpts values, such as normalized Kendall-tau distance or Jaccard distance between the top-n score sets, with ties handled consistently. Then apply Auto-GLOSH's max-orthogonal-distance rule to locate the ORD-Profile elbow and check whether it coincides with the argmax of P@n and with the elbow of the ranking-aware dissimilarity. If the ranking-aware dissimilarity remains high after the ORD-Profile elbow, or if the ORD-Profile elbow does not track P@n on a substantial fraction of held-out datasets, the concern lands: sorted-score dissimilarity is not a valid proxy for ranking stability.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the ORD-Profile elbow marks the minpts where GLOSH's outlier ranking is best (Section IV-C, Definition 4.2). But the ORD-Profile is the Pearson dissimilarity between sorted GLOSH score sequences S_k and S_{k+1}. Sorting removes point identities, making the measure invariant to arbitrary permutations of which point receives which score. P@n, however, depends entirely on whether the true outliers occupy the top n positions, i.e., on the point-to-score mapping. Two consecutive minpts values can have nearly identical sorted score lists while the top-n membership changes completely; the ORD-Profile would report near-zero dissimilarity despite a large change in P@n. Conversely, the ORD-Profile can show structure when only lower-ranked inlier scores change. The paper's own justification in IV-C says the profiles imply 'relative order and magnitude' change minimally, but the measure never examines relative order. Thus the elbow may simply reflect convergence of the marginal score distribution, not ranking stability. This is consistent with the local-outlier failures in Table II: the score distribution can settle while local outliers remain misranked. Since the only evidence for the elbow-to-best-P@n alignment is visual inspection on the motivating Banana dataset (Figures 2 and 3), and since the evaluation includes that same dataset, the central regularity is currently supported mainly by a measure that does not directly test it.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Auto-GLOSH and POLAR, two unsupervised heuristics for removing the main user choices from GLOSH-based outlier detection. Auto-GLOSH computes a per-point GLOSH-Profile over a range of minpts values, then builds an ORD-Profile of Pearson dissimilarities between consecutive sorted GLOSH score lists; the minpts at the elbow of this profile is selected as m*. POLAR then thresholds the GLOSH scores at m* by locating a knee in the sorted score sequence and adjusting it with a linear-regression extrapolation of the inlier trend. The evaluation covers 69 datasets, mostly Banana/Anisotropic/Circular with synthetic global, clump, and local outliers, plus semi-synthetic datasets built from 15 real one-class-classification inlier sets. The reported results show near-oracle P@n for global outliers and clumps, and weaker results for local outliers.","tokens_in":20401,"tokens_out":7053,"duration_ms":67215,"significance":"If the central regularity were established, the contribution would be practically valuable: users could run GLOSH without choosing minpts or a score threshold or a count n, and the CORE-SG construction gives a credible efficiency argument. The dataset count is broad, the comparison against KNN, LOF, ABOD, and COPOD is useful, and the runtime table (Table V) is informative. However, the load-bearing claim that the ORD-Profile elbow coincides with the minpts of best ranking quality is currently supported mainly by a measure that does not directly test ranking quality, and by visual inspection of the same Banana dataset that later appears in the scored tables. The paper is honest about the local-outlier difficulty, but the abstract and conclusions state a stronger, unrestricted claim. The contribution is promising but not yet established.","major_comments":[{"comment":"The ORD-Profile is computed on sorted GLOSH score sequences, which discards the point-to-score mapping. Precision@n depends on whether the true outliers occupy the top n positions, i.e., on exactly that mapping. Two consecutive minpts values can have nearly identical sorted score lists while the top-n membership changes completely; the ORD-Profile would then report near-zero dissimilarity despite a large change in P@n. The text in IV-C states that 'relative order and magnitude' change minimally, but the measure never examines relative order of points. Please provide direct evidence for the claimed correspondence, for example by plotting the Jaccard overlap or rank correlation of the top-n point sets against the ORD-Profile over the same minpts range, across all datasets. This is the central claim of Section IV-C and currently rests on indirect evidence.","section":"IV-C, Definition 4.2 and Eq. (4)"},{"comment":"The central regularity was calibrated on the Banana dataset in Figures 2-4, and the same Banana dataset is included in the main evaluation in Tables II, III, IV, and VI. Since the elbow-to-best-P@n correspondence is asserted from visual inspection of this dataset rather than derived, the near-oracle results on Banana are partly circular. Please report the evaluation on datasets that were not used to motivate the method, or remove Banana from the scored tables, or provide a derivation of the regularity from GLOSH/HDBSCAN* properties.","section":"IV-B and IV-C, Figures 2-4 vs Tables II-IV and VI"},{"comment":"The paper states in IV-B that for local outliers there is 'no notable gap between the profiles of local outliers and inliers' and that one cannot choose an arbitrary value beyond the elbow. Consistently, the local-outlier P@n values in Table II include 0.35 for STAMPS, 0.22 for BREASTW, and 0.76 for Circular, which are far from the claimed best-or-near-best behavior. The abstract and introduction claim general 'best or near best' outlier detection results without restricting to global outliers and clumps. Please either restrict the central claim to the settings where the regularity holds, or provide additional evidence that Auto-GLOSH and POLAR handle local outliers on datasets beyond the current weak cases.","section":"IV-B and Table II, local-outlier columns"},{"comment":"All scored outliers are generated by a single protocol: GMM covariance or mean scaling with alpha=5, uniform global draws, and Tomek-link filtering. The real one-class classification datasets in Table I are explicitly not used for scoring, and the paper argues that their 'labeled outliers' do not behave like statistical outliers. This means the generalization to real outlier structure is unestablished. Please add experiments with different generation settings (varying alpha, noise levels, cluster shapes) or with real OCC datasets under a clearly stated evaluation protocol, and discuss whether the ORD-Profile elbow remains aligned with best P@n when the outlier generation model is changed.","section":"IV-B and VI-A"}],"minor_comments":[{"comment":"The sentence in IV-C that high covariance between consecutive sorted score sequences implies high Pearson dissimilarity is backwards: under Eq. (4), high covariance relative to the variances gives a high correlation and hence a dissimilarity close to 0. Please correct the reasoning or the equation.","section":"IV-C, Eq. (4)"},{"comment":"The description of POLAR says to search for the GLOSH score I that is most similar to the extrapolated score R, but 'most similar' is not defined. Specify the distance used and the behavior when R falls outside the observed score range.","section":"V-B"},{"comment":"The term 'parameter-free' overstates the method: mmax, the i+3 index mapping from the ORD-Profile elbow to minpts, and the linear-regression adjustment are user-specified or heuristic choices. Please qualify the claim, for example by stating that the method removes the need to choose minpts and the outlier count/threshold while retaining a range bound mmax.","section":"Title and Abstract"},{"comment":"The header of Table II contains repeated artifacts such as 'Γ5Γ5Γ5' and 'Γ10Γ10Γ10'; please clean the table formatting so that each column is labeled once.","section":"Table II"},{"comment":"The complexity statement 'assessing the m* value has O(n) time' is imprecise because the ORD-Profile has length O(mmax); please state the complexity consistently in terms of both n and mmax.","section":"IV-E"}],"recommendation":"major_revision","confidential_remarks":"The core idea is interesting and the empirical scope is broad, but the main evidence for the central regularity is currently indirect (sorted-score dissimilarity rather than ranking-quality measures) and partly circular (Banana is both the discovery dataset and a scored dataset). The reversed covariance reasoning in IV-C should also be corrected. I would be willing to reconsider after the authors provide direct evidence that the ORD-Profile elbow tracks top-n ranking quality, and after they either restrict or substantially strengthen the local-outlier claims. Releasing code and data would also help reproducibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The paper proposes two genuinely new heuristics—Auto-GLOSH for choosing HDBSCAN*'s minpts and POLAR for thresholding GLOSH scores—and they seem to work well for global outliers and outlier clumps across a wide range of datasets. The second thing is less flattering: the load-bearing claim that the ORD-Profile elbow marks the minpts where GLOSH ranking quality peaks is not actually supported by the measure. The ORD-Profile is computed on sorted score lists, so it is invariant to which point receives which score, while P@n depends entirely on that mapping. That gap is real, not a nitpick.\n\nWhat is good: the constructs are new, the paper is unusually candid about local outliers being harder, the runtime analysis reusing CORE-SG is sensible, and the evaluation spans 69 datasets. The tables show near-oracle P@n for global and clump outliers on most datasets. Removing two user choices from GLOSH is a practical gain.\n\nSoft spots, in proportion. First, the central regularity is empirical, not derived, and the Banana dataset used to develop the method also appears in the evaluation tables. That is calibration on the test set. Second, the ORD-Profile cannot detect the ranking changes P@n depends on; the paper justifies it by saying relative order and magnitude change minimally, but the measure never examines relative order. The elbow may just reflect convergence of the marginal score distribution. Third, the evaluation is narrow: one outlier generation recipe (alpha=5 GMM scaling plus uniform draws), no error bars or repeated trials, and no released code or data. Fourth, \"parameter-free\" is an overstatement—mmax, the i+3 index offset, and POLAR's regression threshold are all user choices. Finally, local outliers: the abstract promises near-best results, but Table II shows weak local scores like STAMPS 0.35 and BREASTW 0.22. The paper discloses this, but it tempers the headline.\n\nThese are addressable issues, not fatal ones. The heuristic pair is worth a serious referee: I would send it to review and ask for code/data, broader benchmarks, repeated trials, and a rewritten claim that separates the empirical heuristic from a derived guarantee. The right audience is practitioners using GLOSH in unsupervised pipelines and researchers working on automatic parameter selection for density-based detectors. It should not be desk-rejected.","headline":"Useful and novel heuristic pair for removing two user choices from GLOSH, but the central elbow-to-best-P@n link is empirically asserted on the same datasets used to design it, and the ORD-Profile measure never actually looks at point rankings.","tokens_in":21052,"tokens_out":3555,"would_cite":false,"duration_ms":31575,"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":"An elbow in a score curve pins down GLOSH's best settings without any user parameters.","keywords":["outlier detection","GLOSH","HDBSCAN*","minpts selection","automatic threshold","parameter-free","hierarchical density-based clustering","unsupervised learning"],"falsifier":"Take a benchmark dataset with ground-truth outliers, compute the ORD-Profile and GLOSH's Precision@n for every minpts from 2 to 100, and check whether the minpts selected by the maximum-orthogonal-distance rule matches the minpts with maximum Precision@n; any dataset where the two differ by a wide margin, or where the ORD-Profile lacks a distinct elbow, would falsify the central claim. A pointed version: construct a mixture where ranking quality peaks at a high minpts but the curve's elbow falls early, showing the chosen setting underperforms the best achievable setting.","tokens_in":19864,"feed_emoji":"📉","tokens_out":7842,"duration_ms":209755,"temperature":0.7,"pith_summary":"This paper claims that the two parameters a user must otherwise supply to the GLOSH outlier detector — the HDBSCAN* density parameter minpts and the number of outliers to look for — can be chosen automatically with no loss in detection quality. The authors introduce Auto-GLOSH, which builds a per-point profile of GLOSH scores across a range of minpts values and then reads off the value at which consecutive sorted score lists stop changing chaotically and start changing at a uniform rate. That value, found as the elbow of a dissimilarity curve, is claimed to be the minpts where GLOSH achieves its best or near-best precision at finding the true outliers. A second component, POLAR, finds a knee in the sorted score sequence at that minpts and adjusts it by linear regression to set a cutoff that labels potential outliers without knowing how many there are. The payoff is a fully unsupervised, parameter-free version of GLOSH whose reported precision on global, clumped, and mixed outliers matches or approaches the best achievable with any minpts in the tested range.","feed_headline":"Score-curve elbow picks the best GLOSH setting with zero tuning","feed_subtitle":"Auto-GLOSH and POLAR match near-best outlier precision without choosing minpts or outlier count.","key_machinery":"The central objects are the GLOSH–Profile — for each point, the array of its GLOSH scores over minpts values 2 to mmax — and the GLOSH Outlier Rank Dissimilarity–Profile (ORD–Profile), which records 1 minus the Pearson correlation between the sorted GLOSH score sequences at consecutive minpts values. The ORD-Profile's elbow, located by the maximum orthogonal distance from the curve to a chord, is the load-bearing signal: the paper argues that this elbow marks the minpts where rank order and score magnitudes stabilize and where GLOSH's outlier ranking is best. The same maximum-orthogonal-distance geometry, applied to the sorted scores at the chosen minpts and followed by a regression-based adjustment of the knee threshold, carries the threshold-selection half of the argument.","core_discovery":"GLOSH, the outlier detector built into HDBSCAN*, assigns each point a score by comparing its density with the densest point in its cluster in a hierarchical density estimate; the hierarchy, and hence the scores, depend on a user-chosen minpts. The paper's central discovery is an empirical regularity: as minpts increases, the Pearson dissimilarity between the sorted GLOSH score lists of consecutive minpts values is initially large and erratic, then drops to near zero, and the point where it drops — the elbow — coincides with the minpts where GLOSH's Precision@n is best. Auto-GLOSH locates this elbow by maximizing the orthogonal distance from the curve to the straight line joining the curve's maximum and its endpoint, and uses that minpts value. At that value, the sorted GLOSH scores show a knee where outlier scores accelerate away from the inlier trend; POLAR identifies the knee by the same distance rule and then fits a linear regression to the pre-knee inlier scores to set a threshold beyond the knee, so outliers are labelled without a preset count. The paper reports that on 69 datasets spanning global outliers, local outliers, and outlier clumps, this parameter-free pipeline attains the best or near-best GLOSH result on most datasets and is competitive with or better than KNN, LOF, ABOD, and COPOD on the majority of tested cases.","pith_inferences":["Beyond the paper's tests, the elbow rule is a stability-detection heuristic, so an \"elbow strength\" diagnostic could be added: datasets whose ORD-Profile is flat or has multiple elbows would be flagged as unreliable, instead of silently returning a minpts.","The same profile idea may transfer to other parameters of density hierarchies, such as cluster-selection thresholds or mutual-reachability scaling, wherever a stability elbow is expected to mark a regime change in ranking behavior.","On datasets with semantically anomalous but locally dense outliers, the paper's Table I suggests these points look like inliers to several methods; an implicit practical conclusion is that Auto-GLOSH is most reliable when suspected outliers are genuinely density-deviant.","A testable variant would replace the maximum-distance elbow rule with a model-selection rule, such as a piecewise-linear fit of the ORD-Profile, and check whether the chosen minpts tracks best Precision@n more closely on high-dimensional benchmarks."],"forward_implications":["A practitioner can run GLOSH on a new dataset with no parameter choices and obtain Precision@n equal to or close to the best that any minpts in the tested range would give, for global, clumped, and mixed outliers.","The automatically chosen minpts frequently outperforms the common hand-picked values (5, 10, 25, 50, 100), so the method removes a source of sensitivity that prior work identified for GLOSH.","The POLAR threshold gives F-measure and G-Mean close to the best obtainable threshold on most datasets; when recall matters more than precision, using the unadjusted knee threshold is the paper's stated alternative.","Because all hierarchies for the minpts range can be extracted from one core graph (CORE-SG), the entire parameter-free procedure runs in time comparable to a single HDBSCAN* pass rather than mmax separate passes."],"supporting_citations":[{"why":"Defines HDBSCAN*, the mutual reachability hierarchy, and GLOSH, the outlier scoring method whose minpts and threshold this paper automates.","marker":"[5]"},{"why":"CORE-SG shows how to extract MSTs for many minpts values from a single core graph, which keeps the multi-hierarchy construction efficient.","marker":"[8]"},{"why":"Reports that GLOSH performance varies substantially with minpts, motivating the need for automatic selection.","marker":"[13]"},{"why":"Supplies the banana-shaped synthetic datasets used for the motivating profile and elbow visualizations.","marker":"[19]"},{"why":"Provides the real one-class classification datasets and the synthetic outlier generation procedure (global, local, clumps) used in the experiments.","marker":"[20]"},{"why":"Source of the Pearson correlation-based dissimilarity used to compare sorted GLOSH score sequences in the ORD-Profile.","marker":"[23]"},{"why":"Provides the vector parallelogram area formula used to compute orthogonal distances for locating the elbow and knee.","marker":"[24]"},{"why":"Linear regression used by POLAR to estimate the inlier score trend and adjust the threshold beyond the knee.","marker":"[28]"}],"fun_headline_variants":["Auto-GLOSH and POLAR: no tuning, near-best outlier scores","Elbow rule picks GLOSH's minpts, knee sets threshold","Parameter-free GLOSH matches tuned outlier detection","Zero-parameter outlier detection via GLOSH curve elbows","Score elbow auto-tunes GLOSH, knee auto-thresholds"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is empirical rather than proven: the dissimilarity curve between consecutive GLOSH score lists must have a single clean elbow at exactly the density parameter where GLOSH's outlier ranking is best; a real dataset with no such elbow, or with an elbow that does not align with ranking quality, leaves the method without an anchor.","fun_headline_variants_meta":{"raw":{"variants":["Auto-GLOSH and POLAR: no tuning, near-best outlier scores","Elbow rule picks GLOSH's minpts, knee sets threshold","Parameter-free GLOSH matches tuned outlier detection","Zero-parameter outlier detection via GLOSH curve elbows","Score elbow auto-tunes GLOSH, knee auto-thresholds"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000314,"raw_usage":{"total_tokens":1866,"prompt_tokens":1115,"completion_tokens":751,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":731,"completion_tokens_details":{"reasoning_tokens":666}},"tokens_in":731,"tokens_out":751,"duration_ms":6812,"temperature":1.0,"reasoning_tokens":666,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T21:15:39.143483+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a benchmark dataset with ground-truth outliers, compute the ORD-Profile and GLOSH's Precision@n for every minpts from 2 to 100, and check whether the minpts selected by the maximum-orthogonal-distance rule matches the minpts with maximum Precision@n; any dataset where the two differ by a wide margin, or where the ORD-Profile lacks a distinct elbow, would falsify the central claim. A pointed version: construct a mixture where ranking quality peaks at a high minpts but the curve's elbow falls early, showing the chosen setting underperforms the best achievable setting.","supporting_citations":[{"cited_title":"Hierarchical density estimates for data clustering, visualization, and outlier detection,","cited_arxiv_id":null,"evidence_quote":"Defines HDBSCAN*, the mutual reachability hierarchy, and GLOSH, the outlier scoring method whose minpts and threshold this paper automates."},{"cited_title":"Core-sg: Efficient computation of multiple msts for density-based methods,","cited_arxiv_id":null,"evidence_quote":"CORE-SG shows how to extract MSTs for many minpts values from a single core graph, which keeps the multi-hierarchy construction efficient."},{"cited_title":"Swersky, A study of unsupervised outlier detection for one-class classification","cited_arxiv_id":null,"evidence_quote":"Reports that GLOSH performance varies substantially with minpts, motivating the need for automatic selection."},{"cited_title":"Synthetic dataset for outlier detection,","cited_arxiv_id":null,"evidence_quote":"Supplies the banana-shaped synthetic datasets used for the motivating profile and elbow visualizations."},{"cited_title":"Adbench: Anomaly detection benchmark,","cited_arxiv_id":null,"evidence_quote":"Provides the real one-class classification datasets and the synthetic outlier generation procedure (global, local, clumps) used in the experiments."},{"cited_title":"On evaluation of outlier rankings and outlier scores,","cited_arxiv_id":null,"evidence_quote":"Source of the Pearson correlation-based dissimilarity used to compare sorted GLOSH score sequences in the ORD-Profile."},{"cited_title":"Anton and C","cited_arxiv_id":null,"evidence_quote":"Provides the vector parallelogram area formula used to compute orthogonal distances for locating the elbow and knee."},{"cited_title":"Linear regression,","cited_arxiv_id":null,"evidence_quote":"Linear regression used by POLAR to estimate the inlier score trend and adjust the threshold beyond the knee."}],"review_version":1}