Fence-based outlier detectors, Part 1

March 29, 2022 Mathematics Statistics Research Outliers

In previous posts, I discussed properties of Tukey’s fences and asymmetric decile-based outlier detector (Part 1, Part 2). In this post, I discuss the generalization of fence-based outlier detectors.

Notation

A Symmetric Fence-based outlier detector could be defined using the following range:

$$ SF(p, k) = [Q_p - k(Q_{1-p} - Q_p), Q_{1-p} + k(Q_{1-p} - Q_p)] $$

where $Q_s$ is an estimation of the $s^\textrm{th}$ quantile, $p \in [0, 0.5]$.

All the sample elements outside this range are marked as outliers. Using this notation, Tukey’s fences could be defined as $SF(0.25, k)$.

An Asymmetric Fence-based outlier detector is defined using the following range:

$$ AF(p, k) = [Q_p - 2k(Q_{0.5} - Q_{p}), Q_{1-p} + 2k(Q_{1-p} - Q_{0.5})]. $$

An asymmetric decile-based outlier detector could be defined as $AF(0.1, k)$.

Simulation 1

Let’s perform the following experiment:

Enumerate two types of fence-based outlier detectors: asymmetric and symmetric.
Enumerate different $p$ values: $0.1$ (deciles) and $0.25$ (quartiles).
Enumerate different $k$ values: $1.0$, $1.5$, $2.0$, $2.5$, $3.0$, $3.5$, $4.0$.
Enumerate different distributions: the normal distribution, the exponential distribution, the Gumbel distribution.
For each combination of the above parameters, estimate the fence values assuming that $Q_s$ is the true value of $s^\textrm{th}$ quantile. Next, calculate the portion of the distribution outside the fences. Thus, we get the probability of observing a single outlier.

The results are below.

Normal distribution:

type	p	k	outliers
AF	0.10	1.0	0.00012072230941067211
AF	0.10	1.5	0.00000029563871592760
AF	0.10	2.0	0.00000000014767521496
AF	0.10	2.5	0.00000000000001483506
AF	0.10	3.0	0.00000000000000000015
AF	0.10	3.5	0.00000000000000000000
AF	0.10	4.0	0.00000000000000000000
AF	0.25	1.0	0.04302479073838957196
AF	0.25	1.5	0.00697660323928020812
AF	0.25	2.0	0.00074502950319118666
AF	0.25	2.5	0.00005189186759204938
AF	0.25	3.0	0.00000234194246287498
AF	0.25	3.5	0.00000006817407812073
AF	0.25	4.0	0.00000000127585892633
SF	0.10	1.0	0.00012072230941067211
SF	0.10	1.5	0.00000029563871592760
SF	0.10	2.0	0.00000000014767521496
SF	0.10	2.5	0.00000000000001483506
SF	0.10	3.0	0.00000000000000000015
SF	0.10	3.5	0.00000000000000000000
SF	0.10	4.0	0.00000000000000000000
SF	0.25	1.0	0.04302479073838957196
SF	0.25	1.5	0.00697660323928020812
SF	0.25	2.0	0.00074502950319118666
SF	0.25	2.5	0.00005189186759204938
SF	0.25	3.0	0.00000234194246287498
SF	0.25	3.5	0.00000006817407812073
SF	0.25	4.0	0.00000000127585892633

Exponential distribution:

type	p	k	outliers
AF	0.10	1.0	0.00400000000
AF	0.10	1.5	0.00080000000
AF	0.10	2.0	0.00016000000
AF	0.10	2.5	0.00003200000
AF	0.10	3.0	0.00000640000
AF	0.10	3.5	0.00000128000
AF	0.10	4.0	0.00000025600
AF	0.25	1.0	0.06250000000
AF	0.25	1.5	0.03125000000
AF	0.25	2.0	0.01562500000
AF	0.25	2.5	0.00781250000
AF	0.25	3.0	0.00390625000
AF	0.25	3.5	0.00195312500
AF	0.25	4.0	0.00097656250
SF	0.10	1.0	0.01111111111
SF	0.10	1.5	0.00370370370
SF	0.10	2.0	0.00123456790
SF	0.10	2.5	0.00041152263
SF	0.10	3.0	0.00013717421
SF	0.10	3.5	0.00004572474
SF	0.10	4.0	0.00001524158
SF	0.25	1.0	0.08333333333
SF	0.25	1.5	0.04811252243
SF	0.25	2.0	0.02777777778
SF	0.25	2.5	0.01603750748
SF	0.25	3.0	0.00925925926
SF	0.25	3.5	0.00534583583
SF	0.25	4.0	0.00308641975

Gumbel distribution:

type	p	k	outliers
AF	0.10	1.0	0.00243138782251
AF	0.10	1.5	0.00036996004078
AF	0.10	2.0	0.00005624389383
AF	0.10	2.5	0.00000854944973
AF	0.10	3.0	0.00000129954752
AF	0.10	3.5	0.00000019753535
AF	0.10	4.0	0.00000003002599
AF	0.25	1.0	0.05225343350821
AF	0.25	1.5	0.02037237984519
AF	0.25	2.0	0.00849982291839
AF	0.25	2.5	0.00353655486394
AF	0.25	3.0	0.00146932399013
AF	0.25	3.5	0.00061008683006
AF	0.25	4.0	0.00025325410919
SF	0.10	1.0	0.00480943027492
SF	0.10	1.5	0.00103073558090
SF	0.10	2.0	0.00022057403284
SF	0.10	2.5	0.00004718708372
SF	0.10	3.0	0.00001009397656
SF	0.10	3.5	0.00000215921115
SF	0.10	4.0	0.00000046187726
SF	0.25	1.0	0.05920771183234
SF	0.25	1.5	0.02682956730034
SF	0.25	2.0	0.01231232518267
SF	0.25	2.5	0.00562770388698
SF	0.25	3.0	0.00256759594377
SF	0.25	3.5	0.00117046709217
SF	0.25	4.0	0.00053336722084

Simulation 2

Let’s perform the following experiment:

Enumerate two types of fence-based outlier detectors: asymmetric and symmetric.
Enumerate different $p$ values: $0.1$ (deciles) and $0.25$ (quartiles).
Enumerate different $k$ values: $1.0$, $1.5$, $2.0$, $2.5$, $3.0$, $3.5$, $4.0$.
Enumerate different distributions: the normal distribution, the exponential distribution, the Gumbel distribution.
Enumerate different sample sizes $n$: $5$, $10$, $50$, $100$, $500$, $1000$.
For each combination of the above parameters, estimate the fence values assuming that $Q_s$ is the true value of $s^\textrm{th}$ quantile. Next, calculate the probability of having outliers for the given sample size $n$.

The results are below.

Normal distribution, SF, p=0.1:

k/n	5	10	50	100	500	1000
1	0.0006	0.00121	0.00602	0.01200	0.05858	0.11373
1.5	0.0000	0.00000	0.00001	0.00003	0.00015	0.00030
2	0.0000	0.00000	0.00000	0.00000	0.00000	0.00000
2.5	0.0000	0.00000	0.00000	0.00000	0.00000	0.00000
3	0.0000	0.00000	0.00000	0.00000	0.00000	0.00000
3.5	0.0000	0.00000	0.00000	0.00000	0.00000	0.00000
4	0.0000	0.00000	0.00000	0.00000	0.00000	0.00000

Normal distribution, SF, p=0.25:

k/n	5	10	50	100	500	1000
1	0.19739	0.35582	0.88907	0.98770	1.00000	1.00000
1.5	0.03440	0.06762	0.29535	0.50347	0.96982	0.99909
2	0.00372	0.00743	0.03658	0.07182	0.31110	0.52541
2.5	0.00026	0.00052	0.00259	0.00518	0.02561	0.05057
3	0.00001	0.00002	0.00012	0.00023	0.00117	0.00234
3.5	0.00000	0.00000	0.00000	0.00001	0.00003	0.00007
4	0.00000	0.00000	0.00000	0.00000	0.00000	0.00000

Normal distribution, AF, p=0.1:

k/n	5	10	50	100	500	1000
1	0.0006	0.00121	0.00602	0.01200	0.05858	0.11373
1.5	0.0000	0.00000	0.00001	0.00003	0.00015	0.00030
2	0.0000	0.00000	0.00000	0.00000	0.00000	0.00000
2.5	0.0000	0.00000	0.00000	0.00000	0.00000	0.00000
3	0.0000	0.00000	0.00000	0.00000	0.00000	0.00000
3.5	0.0000	0.00000	0.00000	0.00000	0.00000	0.00000
4	0.0000	0.00000	0.00000	0.00000	0.00000	0.00000

Normal distribution, AF, p=0.25:

k/n	5	10	50	100	500	1000
1	0.19739	0.35582	0.88907	0.98770	1.00000	1.00000
1.5	0.03440	0.06762	0.29535	0.50347	0.96982	0.99909
2	0.00372	0.00743	0.03658	0.07182	0.31110	0.52541
2.5	0.00026	0.00052	0.00259	0.00518	0.02561	0.05057
3	0.00001	0.00002	0.00012	0.00023	0.00117	0.00234
3.5	0.00000	0.00000	0.00000	0.00001	0.00003	0.00007
4	0.00000	0.00000	0.00000	0.00000	0.00000	0.00000

Exponential distribution, SF, p=0.1:

k/n	5	10	50	100	500	1000
1	0.05433	0.10572	0.42803	0.67285	0.99625	0.99999
1.5	0.01838	0.03643	0.16934	0.31000	0.84359	0.97554
2	0.00616	0.01228	0.05990	0.11621	0.46080	0.70926
2.5	0.00206	0.00411	0.02037	0.04033	0.18601	0.33742
3	0.00069	0.00137	0.00684	0.01362	0.06629	0.12819
3.5	0.00023	0.00046	0.00228	0.00456	0.02260	0.04470
4	0.00008	0.00015	0.00076	0.00152	0.00759	0.01513

Exponential distribution, SF, p=0.25:

k/n	5	10	50	100	500	1000
1	0.35277	0.58110	0.98710	0.99983	1.00000	1.00000
1.5	0.21850	0.38926	0.91503	0.99278	1.00000	1.00000
2	0.13138	0.24551	0.75550	0.94022	1.00000	1.00000
2.5	0.07766	0.14928	0.55442	0.80146	0.99969	1.00000
3	0.04545	0.08883	0.37194	0.60554	0.99045	0.99991
3.5	0.02644	0.05219	0.23510	0.41493	0.93144	0.99530
4	0.01534	0.03044	0.14321	0.26591	0.78682	0.95455

Exponential distribution, AF, p=0.1:

k/n	5	10	50	100	500	1000
1	0.01984	0.03929	0.18160	0.33022	0.86521	0.98183
1.5	0.00399	0.00797	0.03923	0.07691	0.32979	0.55081
2	0.00080	0.00160	0.00797	0.01587	0.07689	0.14787
2.5	0.00016	0.00032	0.00160	0.00319	0.01587	0.03149
3	0.00003	0.00006	0.00032	0.00064	0.00319	0.00638
3.5	0.00001	0.00001	0.00006	0.00013	0.00064	0.00128
4	0.00000	0.00000	0.00001	0.00003	0.00013	0.00026

Exponential distribution, AF, p=0.25:

k/n	5	10	50	100	500	1000
1	0.27580	0.47554	0.96032	0.99843	1.00000	1.00000
1.5	0.14678	0.27202	0.79555	0.95820	1.00000	1.00000
2	0.07572	0.14571	0.54498	0.79296	0.99962	1.00000
2.5	0.03846	0.07543	0.32440	0.54357	0.98019	0.99961
3	0.01938	0.03838	0.17774	0.32388	0.85871	0.98004
3.5	0.00973	0.01936	0.09313	0.17758	0.62376	0.85844
4	0.00487	0.00972	0.04768	0.09308	0.38647	0.62358

Gumbel distribution, SF, p=0.1:

k/n	5	10	50	100	500	1000
1	0.02382	0.04707	0.21420	0.38252	0.91023	0.99194
1.5	0.00514	0.01026	0.05026	0.09799	0.40288	0.64345
2	0.00110	0.00220	0.01097	0.02182	0.10443	0.19796
2.5	0.00024	0.00047	0.00236	0.00471	0.02332	0.04609
3	0.00005	0.00010	0.00050	0.00101	0.00503	0.01004
3.5	0.00001	0.00002	0.00011	0.00022	0.00108	0.00216
4	0.00000	0.00000	0.00002	0.00005	0.00023	0.00046

Gumbel distribution, SF, p=0.25:

k/n	5	10	50	100	500	1000
1	0.26300	0.45683	0.95272	0.99776	1.00000	1.00000
1.5	0.12714	0.23812	0.74329	0.93410	1.00000	1.00000
2	0.06006	0.11652	0.46175	0.71029	0.99796	1.00000
2.5	0.02782	0.05487	0.24586	0.43128	0.94050	0.99646
3	0.01277	0.02538	0.12063	0.22670	0.72347	0.92353
3.5	0.00584	0.01164	0.05688	0.11052	0.44322	0.68999
4	0.00266	0.00532	0.02632	0.05195	0.23414	0.41346

Gumbel distribution, AF, p=0.1:

k/n	5	10	50	100	500	1000
1	0.01210	0.02405	0.11460	0.21607	0.70393	0.91235
1.5	0.00185	0.00369	0.01833	0.03633	0.16891	0.30929
2	0.00028	0.00056	0.00281	0.00561	0.02773	0.05469
2.5	0.00004	0.00009	0.00043	0.00085	0.00427	0.00851
3	0.00001	0.00001	0.00006	0.00013	0.00065	0.00130
3.5	0.00000	0.00000	0.00001	0.00002	0.00010	0.00020
4	0.00000	0.00000	0.00000	0.00000	0.00002	0.00003

Gumbel distribution, AF, p=0.25:

k/n	5	10	50	100	500	1000
1	0.23535	0.41531	0.93167	0.99533	1.00000	1.00000
1.5	0.09780	0.18603	0.64269	0.87233	0.99997	1.00000
2	0.04178	0.08182	0.34741	0.57413	0.98599	0.99980
2.5	0.01756	0.03481	0.16234	0.29832	0.82991	0.97107
3	0.00733	0.01460	0.07088	0.13674	0.52059	0.77017
3.5	0.00305	0.00608	0.03005	0.05920	0.26298	0.45680
4	0.00127	0.00253	0.01258	0.02501	0.11895	0.22375

References (2)

Asymmetric decile-based outlier detector, Part 1 (2022-03-08) 2 2
Asymmetric decile-based outlier detector, Part 2 (2022-03-15) 2 2

Backlinks (1)

Fence-based outlier detectors, Part 2 (2022-04-05) 1