Towards Efficient Discriminative Pattern Mining in Hybrid Domains

Yoshitaka Kameya

REVIEW 3 major objections 5 minor 30 references

Towards Efficient Discriminative Pattern Mining in Hybrid Domains

T0 review · 3 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Algorithm mines class-distinguishing patterns in hybrid datasets without pre-discretizing numeric attributes.

desk verdict The interval-item concept lattice is a natural extension, but the static merge in §3.1 breaks the exhaustive-covering guarantee, and the experiments don't compare against anything. read the letter →

arxiv 1908.06801 v1 pith:B6UYQQYY submitted 2019-08-15 cs.DB cs.LG

Yoshitaka Kameya This is my paper · ORCID

classification cs.DBcs.LG

keywords discriminativepatternminingemergingsubgroupdiscoveryhybriddomainsintervalitemsconceptlatticeFP-growthexhaustivecovering

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 proposes an algorithm for discriminative pattern mining—finding patterns that distinguish one class from the others—that works directly on hybrid datasets containing both symbolic and numeric attributes. The key idea is to convert each numeric value into a base interval item and to allow patterns to use both base and upper intervals in a concept lattice, then search that space with FP-growth extended by branch-and-bound pruning. The search applies two redundancy constraints, best-covering and closedness-on-the-positives, plus exhaustive covering, which guarantees every positive transaction is covered by some output pattern. The paper demonstrates the algorithm on the iris and German credit datasets, reporting patterns that separate the classes with high F-scores. If correct, the method offers a way to mine discriminative patterns without the usual attribute-wise discretization step.

What carries the argument

The central object is the interval-item concept lattice: for each numeric attribute, cut-points divide the value range into base intervals, and upper intervals are unions of consecutive base intervals, partially ordered by subsumption. The algorithm computes supports of upper intervals from base-interval supports via dynamic programming over this lattice, and prunes the search based on dual-monotonicity of the relevance score, which lets it raise the minimum-support threshold during the search. A static merge operation removes cut-points inside single-class regions, justified by best-covering and dual-monotonicity, and a dynamic merge further combines base intervals that appear only in positive or only in negative conditional transactions, at the cost of an unproven heuristic. The search itself uses exhaustive covering, tracking per-positive-transaction top-1 patterns, and enforces the best-covering and closedness-on-the-positives constraints to cut redundancy.

What would settle it

Construct a small hybrid dataset with two numeric attributes where, for some combination of items, the best F-score pattern requires a cut-point inside a region containing only positive (or only negative) values for that attribute, because the conjunction with an interval on the other attribute makes that cut-point beneficial. Run the proposed algorithm and compare its output to a brute-force enumeration of all interval itemsets; if the optimal pattern is missing or has a lower score than the brute-force best, the merge is unsafe.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that discriminative pattern mining can be extended to hybrid domains by representing numeric attributes as interval items in a concept lattice and searching through all combinations of symbolic, base interval, and upper interval items under two redundancy constraints. The proposed algorithm, built on FP-growth with minimum-support raising and branch-and-bound pruning, uses dual-monotonicity of the relevance score (e.g., F-score) to justify removing cut-points inside regions containing only one class's values, thereby shrinking the search space. It maintains exhaustive covering: for each positive transaction, the search determines its top-1 pattern, and after pruning by the best-covering and closedness-on-the-positives constraints, the union of these per-transaction patterns is returned. The paper reports execution results on the iris dataset and the German credit dataset, showing patterns for each class; for example, the German credit data yields rules such as 'credit amount < 10920 and duration < 66' for the good class.

Load-bearing premise

The search-space reduction assumes that merging cut-points inside single-class regions does not discard any cut-point needed for an optimal pattern; the static merge is justified by dual-monotonicity and the best-covering constraint, but the dynamic merge of base intervals in conditional trees is an additional heuristic without a formal guarantee.

Editorial extensions

If this is right

The algorithm can mine discriminative patterns from hybrid datasets without an attribute-wise discretization step, avoiding the risk of inappropriate discretization.
Exhaustive covering guarantees that every positive transaction is covered by at least one output pattern, so the returned set gives a complete characterization of the target class.
The combination of best-covering and closedness-on-the-positives constraints removes redundant patterns, leaving only those that are locally best for some positive transaction and maximal in positive coverage.
The static merge of cut-points is justified for any dual-monotonic relevance score, so the approach applies not only to F-score but also to χ², information gain, and similar scores.

Reading between the lines

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

The dynamic merge of base intervals in conditional trees lacks a formal correctness guarantee; if it ever discards a cut-point needed by an optimal pattern, the algorithm could miss the best pattern, so the safety of this heuristic is an open question.
The paper notes that runtime is severely affected by the number of base intervals; a natural extension would be to combine this approach with non-greedy histogram-based discretization to reduce the interval count.
Although the paper stops at pattern mining, the mined rule sets could serve as the basis for an associative classifier, and testing classification accuracy on standard benchmarks would give a concrete measure of practical value.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

First, the thing to know: this paper extends your exhaustive covering framework to numeric attributes by introducing interval items over a concept lattice, and the static merge that removes cut-points inside positive-only regions is meant to be justified by dual-monotonicity. That justification is not sound. The claim p(⟨A,[vl,t)⟩∪x|c) < p(⟨A,[vl,v)⟩∪x|c) assumes there is at least one positive transaction in [t,v) that also contains x. If no such transaction exists, the two patterns have identical support and relevance; the more specific one is then the one the closedness constraint should keep. Removing the cut-point deletes it. A simple example: positives at A=10 and 20, negatives at A=0 and 30, cut-points at 5,15,25. Both [5,15) and [15,25) are positive-only and get merged to [5,25). For the pattern {S=a} covering only the A=10 positive, [5,15) and [5,25) have equal support, and only the former is closed. The merge makes the closed pattern unavailable, and the general one is pruned by the closedness test, so that positive transaction ends up with no output pattern. The exhaustive-covering guarantee is therefore relative to the reduced interval language, not the original one. The paper should either prove a version that works with non-strict support differences or explicitly state the guarantee is in the reduced language.

What is genuinely new and worth credit: the concept-lattice representation for interval items, the dynamic programming for computing upper-interval support, and the dynamic merge heuristic in §3.4. The dynamic merge is honestly labeled a heuristic, but it inherits the same closedness problem. The paper is clearly written and the iris patterns are interpretable. The experiments, though, are weak: no baseline comparison at all, and the bad class in German credit takes about seven hours. That does not demonstrate the efficiency claimed in the title.

My overall take: the core idea is plausible and the paper is detailed enough that it deserves a serious referee, but the merge proof needs a fix and the experiments need baselines. If the merge cannot be made lossless, the contribution shrinks to a heuristic without formal guarantees. I would not cite it as a reliable method until that is resolved.

Referee Report

3 major / 5 minor

Summary. The paper proposes an FP-growth-based algorithm for mining discriminative patterns in transactional datasets that contain both symbolic and numeric attributes. Numeric values are first transformed into base interval items using class-informed cut-points, and interval items are organized in a concept lattice so that upper-interval supports can be computed by dynamic programming. The search uses dual-monotonic relevance scores, branch-and-bound pruning with minimum-support raising, and an exhaustive-covering strategy that maintains a top-1 candidate list for each positive transaction; the best-covering and closedness-on-positives constraints are applied to reduce redundancy. The manuscript reports mined patterns for the iris and German credit datasets and claims that the algorithm is efficient for hybrid domains. The central claim is that the proposed algorithm can mine closed, best-covering discriminative patterns while preserving the exhaustive-covering guarantee.

Significance. The idea of representing numeric intervals as items in a concept lattice and computing supports of upper intervals by dynamic programming is natural and potentially useful, and the per-transaction candidate-list formulation of exhaustive covering is clearly specified. The dual-monotonicity-based pruning conditions in Section 2.7 are also derived transparently as a strength of the paper. However, the current manuscript does not establish the advertised guarantees. The static interval-merge step in Section 3.1 is not lossless, the dynamic merge in Section 3.4 is explicitly heuristic, and the experimental section provides no comparison with alternative algorithms and reports a seven-hour runtime on a small dataset. If the merge operations can be made provably safe, or if the claims are explicitly weakened to describe an approximate algorithm, the paper could become a more solid contribution; in its present form, the correctness and efficiency claims are not supported.

major comments (3)

[Section 3.1 (static merge)] The first and third displayed inequalities in Section 3.1 are not guaranteed to be strict. For a positive-only region [u,v) and a cut-point t inside it, one can only conclude p(⟨A,[vl,t)⟩∪x | c) ≤ p(⟨A,[vl,v)⟩∪x | c), with equality whenever no positive transaction satisfying x has an A-value in [t,v). In that case the specific interval and the merged interval have identical positive and negative support relative to the same pattern x, hence identical relevance under F-score; the best-covering constraint does not eliminate the specific pattern, and the closedness-on-positives constraint actively selects the more specific one. Removing t therefore removes the only closed best-covering representative for some positive transactions and invalidates the exhaustive-covering guarantee stated in Section 2.9. Concretely, with transactions t1=(A=10,S=a,c), t2=(A=20,S=b,c), t3=(A=0,S=c,¬c), t4=(A=30,S=d,¬c) and initial cut-points 5, 15, 25, the patterns [5,15)∪{S=a} and [5,25)∪{S=a} tie on positive and negative support, but only the former is closed; after merging [5,15) and [15,25), the closed representative is no longer representable. The analogous argument applies to the negative-only merge.
[Section 3.4 (dynamic merge)] The dynamic merge of base intervals inside conditional FP-trees is introduced as a heuristic, but no correctness argument is given. Because the decision to merge is made on the conditional database for the current prefix x, an interval can appear to be positive-only relative to that prefix while still containing positive transactions that are relevant to some later extension of the pattern. The equality example from Section 3.1 carries over directly: merging two conditional base intervals can delete a cut-point that would be needed for a closed best-covering pattern under the original interval language. As long as this heuristic is employed, the paper cannot claim that the algorithm returns the exact exhaustive-covering, best-covering, closed-on-the-positives answer. The manuscript should either prove that the dynamic merge preserves the output or explicitly characterize the output as an approximation and quantify the loss empirically.
[Section 4 (experiments)] The title and abstract claim efficient discriminative pattern mining, but the experimental section reports no baseline comparison and no quality or scalability evaluation. The German credit run for the bad class visited 2.1 billion candidate patterns and took 24,707 seconds (nearly 7 hours) on a dataset of only 1,000 instances; the iris results in Table 1 are presented without timing or comparison. To support the efficiency claim, the paper needs comparisons with at least one existing numeric subgroup or emerging-pattern miner, such as the methods in [3] or [11], or with a version of the proposed algorithm without the merge heuristics. The runtime also depends on the precision parameter ε and on the number of cut-points, but no values or sensitivity analysis are reported. Without these experiments the central contribution is not demonstrated.

minor comments (5)

[Section 2.1] The precision parameter ε is never given a concrete value or a procedure for setting it; the paper should report the exact ε used and the resulting number of base intervals in the experiments.
[Section 2.4] The notation 'x /nequaly' and 'x /nequaly' appears corrupted; the intended inequality symbols should be typeset properly, for example as x ≠ y.
[Section 4] The abstract mentions both iris and German credit datasets, but Section 4 reports details only for German credit; a table of runtimes and candidate counts for iris would improve completeness.
[Section 3.6] The dynamic re-ordering rule is described only verbally; a pseudocode sketch or a small running example would make the traversal order and the handling of different attribute types clearer.
[Section 5] The concluding remark that the number of base intervals severely affects runtime is important but is not developed into any complexity analysis or practical guidance; a brief complexity statement would help readers understand scalability.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the algorithm's numeric interval handling is evaluated on external benchmarks and does not reduce to its inputs.

full rationale

The paper's derivation chain is an algorithm design rather than a fitted prediction. Numeric values are first converted to base interval items via adaptive cut-points (Section 3.1), then upper interval items are handled through a concept lattice and dynamic programming (Section 3.2), and search is carried out by FP-growth with branch-and-bound pruning, exhaustive covering, and the best-covering/closedness constraints. The central claim that discriminative patterns can be mined in hybrid domains is supported by execution results on the iris and German credit datasets; no quantity is fitted to a subset and then reported as a prediction. The self-citations to the author's prior work ([14] for exhaustive covering and best-covering, [15] for dual-monotonicity) are foundational building blocks, not the target result of this paper. The merge operation in Section 3.1 is justified from dual-monotonicity and the best-covering constraint, and the dynamic merge in Section 3.4 is explicitly described as an observation-based optimization. The skeptical concern about the static merge's strict inequality is a potential soundness/correctness issue, but it is not circularity: the claimed reduction does not make the paper's output equivalent to its input by construction. Since the paper is self-contained against external benchmarks and no load-bearing step reduces to its own assumptions, the appropriate finding is no significant circularity.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The algorithm introduces no new physical or conceptual entities beyond the interval-item representation. Its correctness rests on the dual-monotonicity assumption, a lemma from the author's prior work, the standard FP-growth machinery, and a heuristic dynamic merge that is not formally proven.

free parameters (1)

epsilon (precision for value quantization)
In Section 3.1, an epsilon is pre-defined as a sufficiently small precision to regard all values in [m*epsilon, (m+1)*epsilon) as identical. The choice is left to the user and affects the initial cut-points, hence the search space and results.

assumptions (4)

domain assumption Relevance scores are dual-monotonic with respect to positive and negative support.
Section 2.6 defines dual-monotonicity and uses it to justify branch-and-bound pruning and the merge operation in Section 3.1. The paper states F-score, chi-square, information gain, and support difference are dual-monotonic, but the algorithm's guarantees depend on this property holding for the chosen score.
domain assumption Lemma from [14]: if a pattern y strictly subsumes y' and is more relevant than y', then y' violates the best-covering constraint.
Section 3.1 cites this lemma from the author's own DaWaK-16 paper to justify merging positive-only and negative-only regions. The lemma is not proved in this paper and is load-bearing for the discretization step.
standard math FP-growth is a correct algorithm for frequent pattern mining with FP-trees.
The proposed algorithm is built on FP-growth as described in Section 1 and Section 3, assuming the standard correctness and efficiency properties of FP-trees as established in [12].
ad hoc to paper The dynamic merge of base intervals in conditional FP-trees is safe and does not remove relevant patterns.
Section 3.4 introduces the dynamic merge based on an observation about class imbalance in conditional transactions, but provides no formal proof that this operation preserves the optimal patterns. This is an unverified heuristic.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Towards Efficient Discriminative Pattern Mining in Hybrid Domains." pith.science (2026). https://pith.science/paper/B6UYQQYY

@misc{pith2026190806801,
  author       = {Pith},
  title        = {Pith review of: Towards Efficient Discriminative Pattern Mining in Hybrid Domains},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/B6UYQQYY}},
  note         = {Machine review of arXiv:1908.06801}
}

read the original abstract

Discriminative pattern mining is a data mining task in which we find patterns that distinguish transactions in the class of interest from those in other classes, and is also called emerging pattern mining or subgroup discovery. One practical problem in discriminative pattern mining is how to handle numeric values in the input dataset. In this paper, we propose an algorithm for discriminative pattern mining that can deal with a transactional dataset in a hybrid domain, i.e. the one that includes both symbolic and numeric values. We also show the execution results of a prototype implementation of the proposed algorithm for two standard benchmark datasets.

Figures

Figures reproduced from arXiv: 1908.06801 by the authors.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 30 canonical work pages

[3]

IEEE Trans

Brin, S., Rastogi, R., Shim, K.: Mining optimized gain rul es for numeric attributes. IEEE Trans. on Knowledge and Data Engineering 15(2), 324–338 (2003)

work page 2003
[11]

Data Mining and Knowledge Discovery 19(2), 210–226 (2009)

Grosskreutz, H., R¨ uping, S.: On subgroup discovery in numerical domains. Data Mining and Knowledge Discovery 19(2), 210–226 (2009)

work page 2009
[1]

In: Proc

Atzmueller, M., Lemmerich, F.: Fast subgroup discovery f or continuous target concepts. In: Proc. of ISMIS-09. pp. 35 –44 (2009)

work page 2009
[2]

Data Mining and Knowl- edge Discovery 4, 217–240 (2000)

Bayardo, R., Agrawal, R., Gunopulos, D.: Constraint-bas ed rule mining in large, dense databases. Data Mining and Knowl- edge Discovery 4, 217–240 (2000)

work page 2000
[4]

In: Proc

Domingos, P .: The RISE system: conquering without separa ting. In: Proc. of ICTAI-94. pp. 704–707 (1994)

work page 1994
[5]

(eds.): Contrast Data Mining: Concep ts, Algorithms, and Applications

Dong, G., Bailey, J. (eds.): Contrast Data Mining: Concep ts, Algorithms, and Applications. CRC Press (2012)

work page 2012
[6]

In: Proc

Fayyad, U.M., Irani, K.B.: Multi-interval discretizati on of continuous-valued attributes for classiﬁcation lear ning. In: Proc. of IJCAI-93. pp. 1022–1029 (1993)

work page 1993
[7]

Statistics and Computing 9, 123–143 (1999)

Friedman, J.H., Fisher, N.I.: Bump hunting in high-dimen sional data. Statistics and Computing 9, 123–143 (1999)

work page 1999

Show all 30 references

[8]

I n: Proc

Fukuda, T., Morimoto, Y ., Morishita, S., Tokuyama, T.: Mi ning optimized association rules for numeric attributes. I n: Proc. of PODS-96. pp. 182–191 (1996)

1996
[9]

Springer (2012)

F¨ urnkranz, J., Gamberger, D., Lavraˇ c, N.: Foundationsof Rule Learning. Springer (2012)

2012
[10]

Garriga, G.C., Kralj, P ., Lavraˇ c, N.: Closed sets for labeled data. J. of Machine Learning Research 9, 559–580 (2008 )

2008
[12]

In: Proc

Han, J., Pei, J., Yin, Y .: Mining frequent patterns witho ut candidate generation. In: Proc. of SIGMOD-00. pp. 1–12 (2 000)
[13]

In: Proc

Han, J., Wang, J., Lu, Y ., Tzvetkov, P .: Mining top- k frequent closed patterns without minimum support. In: Proc . of ICDM-02. pp. 211–218 (2002)

2002
[14]

I n: Proc

Kameya, Y .: An exhaustive covering approach to paramete r-free mining of non-redundant discriminative itemsets. I n: Proc. of DaWaK-16. pp. 143–159 (2016)

2016
[15]

In: Proc

Kameya, Y ., Asaoka, H.: Depth-ﬁrst traversal over a mirr ored space for non-redundant discriminative itemsets. In: Proc. of DaWaK-13. pp. 196–208 (2013)

2013
[16]

In: Proc

Kameya, Y ., Ito, K.: Dynamic re-ordering in mining top- k productive discriminative patterns. In: Proc. of TAAI-17. pp. 172–177 (2017)

2017
[17]

I n: Proc

Kameya, Y ., Sato, T.: RP-growth: top- k mining of relevant patterns with minimum support raising. I n: Proc. of SDM-12. pp. 816–827 (2012)

2012
[18]

In: Proc

Kontkanen, P ., Myllym¨ aki, P .: MDL histogram density estimation. In: Proc. of AISTA TS-07. pp. 219–226 (2007)

2007
[19]

Kralj Novak, P ., Lavraˇ c, N., Webb, G.I.: Supervised descriptive rule discovery: a unifying survey of contrast set, emerging pattern and subgroup mining. J. of Machine Learning Researc h 10, 377–403 (2009)

2009
[20]

In: Proc

Morishita, S., Sese, J.: Traversing itemset lattices wi th statistical metric pruning. In: Proc. of PODS-00. pp. 226 –236 (2000)

2000
[21]

In: Proc

Pasquier, N., Bastide, Y ., Taouli, R., Lakhal, L.: Disco vering frequent closed itemsets for association rules. In: Proc. of ICDT-99. pp. 398–416 (1999)

1999
[22]

In: Knowledge Discovery in Databas es, pp

Piatetsky-Shapiro, G.: Discovery, analysis, and prese ntation of strong rules. In: Knowledge Discovery in Databas es, pp. 229–248. AAAI Press (1991)

1991
[23]

In: Proc

Soulet, A., Cr´ emilleux, B., Rioult, F.: Condensed repr esentation of emerging patterns. In: Proc. of PAKDD-04. pp. 127– 132 (2004)

2004
[24]

Knowledge Engineering Review 22(1), 37–65 (2007)

Thabtah, F.: A review of associative classiﬁcation mini ng. Knowledge Engineering Review 22(1), 37–65 (2007)

2007
[25]

In: Proc

Uno, T., Asai, T., Uchida, Y ., Arimura, H.: An eﬃcient algorithm for enumerating closed patterns in transac tion databases. In: Proc. of DS-04. pp. 16–31 (2004)

2004
[26]

In: Proc

Wang, J., Karypis, G.: HARMONY: e ﬃciently mining the best rules for classiﬁcation. In: Proc. o f SDM-05. pp. 205–216 (2005)

2005
[27]

Webb, G.I.: OPUS: an e ﬀﬁcient admissible algorithm for unordered search. J. of Art iﬁcial Intelligence Research 3, 431– 465 (1995)

1995
[28]

Machine L earning 68, 1–33 (2007)

Webb, G.I.: Discovering signiﬁcant patterns. Machine L earning 68, 1–33 (2007)

2007
[29]

Morgan Kaufmann, 2nd edn

Witten, I.H., Frank, E.: Data Mining: Practical Machine Learning Tools and Techniques. Morgan Kaufmann, 2nd edn. (2005)

2005
[30]

Machine Learning 77, 125–15 9 (2009) 8 T able 2.Discriminative patterns mined by the proposed algorithm fr om the german credit dataset

Zimmermann, A., De Raedt, L.: Cluster grouping: from sub group discovery to clustering. Machine Learning 77, 125–15 9 (2009) 8 T able 2.Discriminative patterns mined by the proposed algorithm fr om the german credit dataset. Class c p(c | x) p(x | c) F c(x) Pattern x good 0.71...

2009

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [3]

IEEE Trans

Brin, S., Rastogi, R., Shim, K.: Mining optimized gain rul es for numeric attributes. IEEE Trans. on Knowledge and Data Engineering 15(2), 324–338 (2003)

work page 2003

[2] [11]

Data Mining and Knowledge Discovery 19(2), 210–226 (2009)

Grosskreutz, H., R¨ uping, S.: On subgroup discovery in numerical domains. Data Mining and Knowledge Discovery 19(2), 210–226 (2009)

work page 2009

[3] [1]

In: Proc

Atzmueller, M., Lemmerich, F.: Fast subgroup discovery f or continuous target concepts. In: Proc. of ISMIS-09. pp. 35 –44 (2009)

work page 2009

[4] [2]

Data Mining and Knowl- edge Discovery 4, 217–240 (2000)

Bayardo, R., Agrawal, R., Gunopulos, D.: Constraint-bas ed rule mining in large, dense databases. Data Mining and Knowl- edge Discovery 4, 217–240 (2000)

work page 2000

[5] [4]

In: Proc

Domingos, P .: The RISE system: conquering without separa ting. In: Proc. of ICTAI-94. pp. 704–707 (1994)

work page 1994

[6] [5]

(eds.): Contrast Data Mining: Concep ts, Algorithms, and Applications

Dong, G., Bailey, J. (eds.): Contrast Data Mining: Concep ts, Algorithms, and Applications. CRC Press (2012)

work page 2012

[7] [6]

In: Proc

Fayyad, U.M., Irani, K.B.: Multi-interval discretizati on of continuous-valued attributes for classiﬁcation lear ning. In: Proc. of IJCAI-93. pp. 1022–1029 (1993)

work page 1993

[8] [7]

Statistics and Computing 9, 123–143 (1999)

Friedman, J.H., Fisher, N.I.: Bump hunting in high-dimen sional data. Statistics and Computing 9, 123–143 (1999)

work page 1999