Statistical selection of relevant subspace projections for outlier ranking

Abstract

Outlier mining is an important data analysis task to distinguish exceptional outliers from regular objects. For outlier mining in the full data space, there are well established methods which are successful in measuring the degree of deviation for outlier ranking. However, in recent applications traditional outlier mining approaches miss outliers as they are hidden in subspace projections. Especially, outlier ranking approaches measuring deviation on all available attributes miss outliers deviating from their local neighborhood only in subsets of the attributes. In this work, we propose a novel outlier ranking based on the objects deviation in a statistically selected set of relevant subspace projections. This ensures to find objects deviating in multiple relevant subspaces, while it excludes irrelevant projections showing no clear contrast between outliers and the residual objects. Thus, we tackle the general challenges of detecting outliers hidden in subspaces of the data. We provide a selection of subspaces with high contrast and propose a novel ranking based on an adaptive degree of deviation in arbitrary subspaces. In thorough experiments on real and synthetic data we show that our approach outperforms competing outlier ranking approaches by detecting outliers in arbitrary subspace projections.

Full text

Page 1

Statistical Selection of Relevant Subspace
Projections for Outlier Ranking
Emmanuel M¨ uller•, Matthias Schiffer◦, Thomas Seidl◦
•Karlsruhe Institute of Technology (KIT), Germany
emmanuel.mueller@kit.edu
◦RWTH Aachen University, Germany
{mschiffer, seidl}@cs.rwth-aachen.de
Abstract—Outlier mining is an important data analysis task to
distinguish exceptional outliers from regular objects. For outlier
mining in the full data space, there are well established methods
which are successful in measuring the degree of deviation for out-
lier ranking. However, in recent applications traditional outlier
mining approaches miss outliers as they are hidden in subspace
projections. Especially, outlier ranking approaches measuring
deviation on all available attributes miss outliers deviating from
their local neighborhood only in subsets of the attributes.
In this work, we propose a novel outlier ranking based on
the objects deviation in a statistically selected set of relevant
subspace projections. This ensures to find objects deviating in
multiple relevant subspaces, while it excludes irrelevant projec-
tions showing no clear contrast between outliers and the residual
objects. Thus, we tackle the general challenges of detecting
outliers hidden in subspaces of the data. We provide a selection
of subspaces with high contrast and propose a novel ranking
based on an adaptive degree of deviation in arbitrary subspaces.
In thorough experiments on real and synthetic data we show that
our approach outperforms competing outlier ranking approaches
by detecting outliers in arbitrary subspace projections.
I. INTRODUCTION
Outlier mining has become an important data mining task to
detect inconsistent or suspicious objects in large databases. For
recent applications, outlier mining as an unsupervised learning
task is important for consistency checks of sensor network
measurements, fraud detection in financial transactions, emer-
gency detection in health surveillance and many more. As
measuring and storing of data has become very cheap, in all
of these applications, objects are described by many attributes.
However, for each object only subsets of relevant attributes
provide the meaningful information, the residual attributes are
irrelevant for this object. For example in health surveillance,
for one patient attributes such as “age” and “skin humidity”
might be important to detect the abnormal “dehydration” status
of this patient. Other attributes such as “heart beat rate” are
irrelevant for the detection of this outlier, but are relevant for
the detection of abnormal patients with a heart disease. All
of these attributes are required for some outlier detection, but
each outlier occurs only in subsets of these attributes. Thus,
the distinction between outliers and regular objects is heavily
hindered by considering all available attributes as typically
done in traditional outlier mining methods.
Traditional techniques are well established for outlier min-
ing in the full space, but miss outliers which are hidden in
subspace projections. Thus, our general aim is to develop a
novel outlier ranking based on object deviation in subspace
projections. We focus on outlier ranking a special research
field of outlier mining which sorts objects according to their
local degree of deviation. Local outlier rankings use the local
neighborhood around each object to report an ordered list
presenting the most outlying object first. They provide more
information than just the binary decision about being an
outlier or not. Outlier rankings provide for each object the
extent of outlierness. However, traditional outlier rankings
using outlierness measures in full space are not appropriate
for outliers hidden in subspaces. In the full space all objects
appear to be alike so that traditional outlier rankings cannot
distinguish the outlierness of objects any more.
In this work we measure outlierness (the degree of devia-
tion) of an object in projections of the database taking only
subsets of attributes into account. Hence, we can successfully
detect an outlier in a set of relevant subspace projections in
which this outlier stands out from its surrounding objects. We
model various behaviors of one object in different projections
of the database. An object might show high deviation com-
pared to its neighborhood in one subspace. In addition, the
same object might cluster with some other objects in a second
subspace or might not show up as an outlier in a third scattered
subspace where all objects seem to be outliers. To illustrate
this, we have depicted several projections of a toy example
with two hidden outliers in Figure 1. Please note, that each
object requires the detection of individual subspace projections
to detect its outlier properties. This is in contrast to related
paradigms such as subspace clustering [14] or dimensionality
reduction techniques [10]. Global dimensionality reduction
techniques such as principal components analysis provide only
a single projection for all objects. In contrast, we aim at
detecting multiple relevant subspaces per object. Subspace
clustering detects multiple projections, however, focuses on
subspaces for groups of clustered objects. For outlier ranking,
the focus is on the individual objects and subspaces in which
an outlier is highly deviating from its local neighborhood.
Thus, outlier ranking in subspaces poses novel challenges not
tackled by these research topics.

Page 2

subspaces {1} {2} subspace {1,2,3}subspace {3,4} subspace {1,2}
dense regions
no outliers
dense regions
one outlier
dense regions
another outlier
scattered space
all seem outliers
Fig. 1.Example: Outliers in arbitrary subspaces
In this work, we focus on two key properties for outlier
ranking in subspace projections. First, for each object we
statistically select a set of projections for outlier ranking,
we call these relevant subspaces. In relevant subspaces the
neighborhood of an object is clustered and the object is an
outlier if it deviates from these clustered objects. In case
these relevant subspaces inhabit different clusters in terms
of objects, some objects might be members of a cluster in
one subspace while being outliers in others at the same time.
In contrast, in irrelevant subspaces the neighborhood of an
object is distributed uniformly random such that all objects
seem to be outliers. Overall, our outlier ranking is confronted
with arbitrary subspaces, while only very few are relevant
and contribute to a distinction between clustered objects and
outliers. Second, object deviation increases with the number of
attributes in a relevant subspace. As distances between objects
grow more and more alike due to the “curse of dimensionality”
[5], objects are clustered in dense regions in one dimensional
subspaces while objects are scattered in higher dimensional
spaces (cf. Figure 1). In general, the deviation of objects is
highly influenced by the number of attributes in the considered
subspaces. Thus, for outlier ranking in subspace projections,
we have to cope with two major challenges:
• Outliers appear only in relevant subspaces.
• Incomparable deviation in different subspaces.
To tackle both of these challenges, we propose OUTRES a
new method for outlier ranking in relevant subspace projec-
tions. For our outlier ranking we consider only a selection of
non-uniformly distributed projections. We exclude uniformly
random distributed subspaces by a statistical test, as they
hinder the distinction between outliers and regular objects.
Furthermore, we propose a novel adaptive outlier ranking
measuring comparable degrees of deviation for objects in
arbitrary subspaces. We define outliers to be objects highly
deviating from the estimated density in their local subspace
neighborhood. Overall in contrast to traditional outlier rank-
ing approaches, our outlier ranking considers deviation in
subspaces and adapts to the number of attributes in each
considered subspace.
II. RELATED WORK
In general, outliers are objects that deviate from the rest of
the data to a great extent. However, there have been various
outlier models proposed in the literature. We categorize these
models into two paradigms, traditional outlier mining methods
and subspace outlier mining techniques.
a) Traditional Methods:
Different models have been proposed modeling deviation glob-
ally e.g. in distance-based [12], cluster-based [9] or statistical-
based [4] outlier mining methods. However, such techniques
suffer from difficulties in parametrization, as the extent of
deviation is usually hard to quantify globally. This has led to
outlier ranking based on the local degree of deviation for each
object, as in the well established local outlier factor (LOF)
approach [6] or its extension (LOCI) based on local deviation
[24]. Further extensions have been proposed based on this
general local outlier factor idea. The most recent approach
proposes an angle based outlier factor (ABOF) [15]. Based on
the assumption, that angles between objects are more stable
than distances, ABOF computes for each object an angle range
to the residual objects. However, all of these outlier ranking
methods base on the full space, and thus, fail to separate
outliers from regular objects in subspace projections.
b) Subspace Methods:
In contrast, recent approaches consider subspace projections
for outlier ranking. The key property for all of these ap-
proaches is the appropriate choice of considered subspaces.
As the most basic approach a random choice of subspaces has
been proposed for outlier ranking in subspaces by RPLOF
[16] and Isolation Forest [17]. However, clearly such simple
heuristics might miss outliers due to the random selection
of subspaces. Recently, a more meaningful selection has
been proposed that selects only one subspace spanned as a
hyperplane by a set of reference points (SOD) [13]. Its general
hypothesis states that outliers deviate within this hyperplane.
However, SOD determines the outlierness of an object only
in this single subspace, if objects deviate in two or more
subspaces SOD is unable to distinguish between their outlier
factors.

Page 3

Recent approaches base on subspace clustering as a related
paradigm. As first approach one based on the subspace clus-
tering method CLIQUE [2] to derive objects deviating from
subspace clusters [1]. However, designed as a binary decision
this approach does not provide an outlier ranking. As first
outlier ranking based on subspace clusters a ranking function
using cluster properties as indicators for outliers has been
proposed (OutRank) [20]. However, based on the aggregated
information of clusters one ignores the actual deviation of each
object in the considered subspaces.
Overall, traditional full space outlier ranking methods sim-
ply compute object deviation in one fixed space and thus miss
outliers in subspace projections. In contrast, outlier ranking in
subspace projections take arbitrary projections into account.
However, none of the proposed subspace methods considers a
meaningful selection of relevant subspaces. Furthermore, they
all ignore the incomparable deviation of objects in different
subspaces for their outlierness measures.
III. OUTLIER RANKING IN SUBSPACES
Our general idea is to measure deviation of each object in a
set of relevant subspace projections. In contrast to traditional
outlier ranking approaches, we consider for each object its
deviation in multiple subspaces. This ensures to find objects
deviating in projections of the data, but, it also poses new
major challenges for outlier ranking. In the following we
start with some basic notions and provide a formalization
of these challenges before we propose our novel selection
of significantly non-uniformly distributed subspaces and our
novel adaptive outlierness measure.
A. Notions and Challenges
In general, the aim of outlier ranking is to provide a sorting
of all objects o given in a database DB. Technically, one ranks
according to the degree of deviation measured by a ranking
function r : DB → R. The ranking function provides a real
valued measure of the objects’ outlierness. Ranking functions
can be defined arbitrarily based on the object’s features o =
(o1,...,od). In contrast to traditional approaches that measure
the degree of deviation in the full d-dimensional space D =
{1,...,d}, we measure deviation in subspace projections S ⊆
D. Thus, we ensure to find outliers hidden in any possible
subspace projection.
The general challenge for outlier ranking approaches, is
to provide a meaningful ranking function which achieves to
distinguish between an outlier object o and a regular object
p by providing a clear distinction: r(o) ? r(p). However,
traditional outlier ranking functions fail for outliers hidden in
subspaces as they provide for all objects very similar ranking
values r(o) ≈ r(p) ∀o,p ∈ DB. This can be explained
by the scattered full space of such data sets. While each
object has only a subset of relevant attributes, the residual
attributes provide more or less random values. Considering
distances between objects using all of these attributes one
observes an effect termed the “curse of dimensionality”. As
traditional ranking functions consider all attributes for distance
computation distD(o,p) they cannot distinguish outliers from
regular objects. For the typically used Euclidean distance
??
, distances between o ∈ DB and any residual objects grow
more and more alike with increasing number of attributes
|D| → ∞:
distD(o,p) =
i∈D
(oi− pi)2
lim
|D|→∞
maxp∈DBdistD(o,p) − minp∈DBdistD(o,p)
minp∈DBdistD(o,p)
→ 0
As consequence, ranking values based on these full space
distances become meaningless:
lim
|D|→∞r(o) − r(p) → 0 ⇒ r(o) ≈ r(p) ∀o,p ∈ DB
Although outliers do not show up in full space, they deviate
in subspace projections. Thus, we cope with the curse of
dimensionality by considering the outlierness of each object
in a selection of relevant subspaces. This set of relevant
subspaces RS(o) is selected individually for each object o
such that these subspaces provide a high contrast between o
and its surrounding neighborhood. We measure the outlierness
score(o,S) by restricting distance functions distS(o,p) to the
subspace dimensions in S. The overall ranking value r(o) of an
object o is then simply computed by aggregating its outlierness
in all relevant subspaces:
Definition 1: Subspace Ranking Function
The overall ranking value r(o) of an object o ∈ DB w.r.t. a
set of relevant subspaces RS(o) and an outlierness measure
score(o,S) is defined as:
?
r(o) =
S∈RS(o)
score(o,S)
As aggregation of all outlierness measures in different
subspaces one could use several meaningful functions. Please
note, that we use scoring values in the range of 0...1
with outliers represented by low scores. Thus, the minimum
over all scorings would provide a meaningful aggregation.
However, this would highlight only the outlierness in one
single subspace. Using the sum of scores as aggregation would
lead to low contrast as objects found in clusters with high score
values would blur the overall ranking value. In contrast, we
use the product incorporating outlier properties from different
subspaces such that low scores in multiple subspaces highlight
an object as clear outlier providing high contrast between
outliers and regular objects.
While traditional ranking functions consider the outlierness
of an object only in the full space D, we aim at considering
outlierness in a set of subspaces RS(o) ⊆ P(D) out of the
powerset of possible subspace projections. This is meaningful

Page 4

as outliers might be hidden in multiple subspace projections.
However, two novel challenges arise:
• How to choose the set of relevant subspaces RS(o) for
meaningful outlier ranking
• How
score(o,S) over multiple subspaces S ∈ RS(o).
To tackle these challenges, our key hypothesis is that
outliers can be distinguished in local neighborhoods of non-
uniformly distributed subspaces. We base on the idea of local
outlier ranking as proposed by full space approaches [6], [24].
According to this outlier mining paradigm, we define outliers
as objects that are highly deviating from their local neighbor-
hood. Traditional methods already observed varying density
distributions and proposed local outlierness measures to tackle
outlier ranking in full space. For subspace outlier ranking
we observe an additional factor for varying densities not yet
addressed in the literature. Different subspace provide highly
varying densities ranging from densely clustered subspaces up
to uniformly distributed subspace. To tackle this variance in
densities we propose a statistical selection of subspaces and
our novel adaptive outlierness measure for outlier ranking in
relevant subspaces.
As a density-based approach OUTRES measures its out-
lierness score(o,S) according to the density den(o,S) of
an object in subspace S. Low density values on an object
indicate its outlierness and lead to low scores. However,
objects might be outliers in multiple subspaces, thus, a mean-
ingful outlierness measure has to be comparable over different
subspaces. We will propose an instantiation of our outlierness
function score(o,S) in Section III-C, where we will also give
details about the underlying adaptive density den(o,S) for a
comparable outlierness measurement. As a flexible approach
any density estimation method could be used in our outlierness
measure. Thus, for the general discussion in this section let
us assume the typically used density in a local neighborhood
defined as the number of objects in ε-distance to the object o
in subspace S:
to achieve comparable outliernessvalues
den(o,S) = |N(o,S)| = |{p | distS(o,p) ≤ ε}|
Based on this common density instantiation one can formally
derive two major challenges for an outlier ranking in sub-
spaces:
Challenge 1: Comparability of Outlierness
Outlierness measures are not comparable over multiple sub-
spaces if: For subspace S,T ⊆ D with T ⊂ S
due to curse of dimensionality
⇒
⇒
⇒
⇒
As density drops for increasing dimensionality, outlierness
measures based on density in subspace projections are biased
w.r.t. the dimensionality of the considered subspaces. As stated
in Challenge 1, considering a subspace T and one of its higher
∀p ∈ DB : distS(o,p) ≥ distT(o,p)
den(o,S) ≤ den(o,T)
score(o,S) ≤ score(o,T)
outlierness is biased w.r.t. dimensionality
dimensional projections S, density drops from T to S. Thus,
overall aggregation (cf. Def. 1) of outlierness is hindered by
incomparable measures. Please keep in mind that we set low
values in score(o,S) for highly deviating objects as we sort
our ranking in ascending order. With such an incomparable
measure, subspaces with many attributes would dominate the
ranking value and outliers in low dimensional projections
could not show up in the overall ranking. Thus, as we take
multiple subspaces into account we have to provide an adaptive
outlierness measure with comparable outlierness in arbitrary
subspace projections to achieve a fair ranking of objects in
any subspace.
Challenge 2: Relevance of Subspaces
A subspace S hinders the distinction of outliers if:
S is distributed uniformly random
∀o,p,q ∈ DB : distS(o,q) ≈ distS(p,q)
∀o,p ∈ DB : den(o,S) ≈ den(p,S)
∀o,p ∈ DB : score(o,S) ≈ score(p,S)
distinction of outliers is hindered
⇒
⇒
⇒
⇒
Obviously the full space D is such an irrelevant subspace for
increasing number of attributes |D| → ∞
With decreasing density, one reaches subspaces with uni-
formly distributed objects where outliers do not show up. In-
cluding such an irrelevant subspace projection S into a ranking
function yields very similar ranking values for all objects.
Thus, our key property for the set of relevant subspaces is
to exclude subspaces which are distributed uniformly random.
B. Selection of Relevant Subspaces
First, we propose a statistical selection of the set of relevant
subspaces RS(o) that can distinguish between the object o
and its local neighborhoods in the selected subspaces. As
motivated by Challenge 2, such a distinction based on the
objects density is not possible in scattered subspaces that show
uniformly random distributed data due to the low contrast
between outliers and regular objects. Thus, we propose to
exclude such scattered subspaces from outlier ranking by
testing the underlying distribution in the local neighborhood
N(o,S).
Our test is based on a statistical significance test aiming at
reducing the probability that a uniformly distributed subspace
passes into the set of relevant subspaces. We test against
the null hypothesis that data is uniformly distributed with
|N(o,S)| ∼ Binomial(|DB|,vol(N(o,S))). W.l.o.g, we
assume that data is normalized to 0...1 such that the volume
of N(o,S) provides us the probability of observing one object
in this neighborhood. As given for uniformly distributed data,
the expected number of objects is then |DB| · vol(N(o,S)).
Each object has equal probability of being in the neighborhood
depending only on the neighborhoods volume. Based on this
null hypothesis, we define H0(S is irrelevant) and H1(S is
a relevant subspace) for object o. As uniformly distributed
subspaces hinder the detection of meaningful outliers, this
definition ensures with a given significance level α that

Page 5

density
subspace dimensionality
1234
dense
subspaces
scattered
subspaces
subspace
subspace
subspace 2,3 2,3
1,2
1,4
11443322
1,21,2 1,41,42,3 2,3 3,43,4
1,2,3 1,2,3 2,3,42,3,4 1,3,41,3,41,2,41,2,4
1,2,3,4 1,2,3,4
1,3
2,4
scattered
subspaces
overall dense
subspaces
dimension 3
is irrelevant
relevant
subspaces
Fig. 2. Relevant subspace projections for outlier mining
uniformly distributed subspaces are only included into the
ranking with a very low probability of less than α. While for
uniformly distributed data one expects |DB| · vol(N(o,S))
many objects in the neighborhood, a relevant subspace should
contain significantly more objects.
Definition 2: Relevance Test
For subspace S and neighborhood N(o,S) we define hypothe-
ses H0and H1:
H0:
H1:
S is distributed uniformly random in N(o,S)
S is distributed non-uniformly in N(o,S)
ensuring significantly low first error:
P(H0is rejected |H0is true ) ≤ α
As statistical tool for testing uniform distribution we use the
Kolmogorov-Smirnov goodness of fit test for the uniform dis-
tribution [28]. In recent mining tasks this test has shown good
performance for subspace cluster detection [18]. In contrast to
subspace clustering where one is interested in sets of objects
grouped in a certain subspace, we select relevant subspaces for
each object to distinguish between outliers and regular objects.
For good ranking quality, we have to ensure that such uniform
subspace are only included in very rare cases by setting a low
α value. As significance level for the statistical hypothesis test
we set α = 0.01. Thus, the probability of wrongly rejecting
the hypothesis H0 (the subspace is uniformly distributed) is
only 1%, i.e. for one out of hundred uniform subspaces the
test will make an error and state that this subspace is relevant.
We will show the influence of the α parameter for the overall
outlier ranking quality in Section IV.
As illustrated on the left side of Figure 2, we exclude
uniformly distributed subspaces for each object o individually.
Starting considering 1d projections first, typically these low
dimensional projects are uniformly dense. The whole database
seems to be one dense region. Furthermore, outliers in 1d
projections could be easily detected as pre-processing. By
including more and more dimensions, due to correlations
of the data, the database diverts in multiple dense regions.
Density shows a high variance between dense clusters and
deviating outliers. For our outlier ranking we only take the
outlierness of objects in these relevant subspaces into account.
Adding even more dimensions the subspaces become scattered
like the full space. All objects seem to be outliers.
For a toy example we have depicted three subspaces where
a single hidden outlier is clearly deviating only in the relevant
subspace {1,2}. Measuring outlierness in this subspace yields
a clear distinction between this outlier and its local neigh-
borhood. Please note, it is crucial to exclude the scattered
subspace {1,4} as all objects seem outliers. Thus, outlierness
measures would lose their contrast as low score values are
provided for all objects. In contrast, subspace {2,3} is not
excluded by the relevance test as dense regions with high
scores do not affect the contrast of Def. 1.
Our general idea is to include only subspaces which are
distributed significantly different then the uniformly random
distribution. Hence, we exclude the subspaces that do not
provide any distinction between objects and hinder our outlier
detection. A key observation for relevant subspaces is that
with increasing number of attributes in a subspace S, one
reaches subspaces with uniformly distributed objects where
outliers do not show up any more. By including more and more
attributes distances between objects grow more and more alike
[5]. Thus, the selection of relevant subspaces can be reduced to
the selection of significant attributes to be included in a given
subspace projection S, as stated in the following corrollary:
Corrollary 1: Uniformly distributed subspaces
Let S = {d1,...,dk} be a subspace. Then it holds true:
S uniformly distributed ⇒
d1uniformly distributed
∧ ... ∧
dkuniformly distributed
Consequently, by testing each attribute diwe can assure that
no uniformly distributed subspace is included in the set of
relevant subspaces RS(o). Moreover, we discard a subspace
based on these insights as soon as at least one attribute is
distributed uniformly random.
We base on statistical tests to detect significant subspaces
by excluding uniformly distributed attributes from further
consideration. We perform an incremental processing of the
subspaces including in each step an additional attribute for the

Page 6

considered subspace S. By adding attribute dito S we check
if objects are uniformly distributed in di. We call an attribute
direlevant for outlier ranking, if objects are significantly non-
uniformly distributed. In contrast to other attributes, a relevant
attribute might be added to S while preserving the clustered
regions of subspace S also in subspace S∪di. Summing up, we
detect meaningful outliers by searching subspaces consisting
only of relevant attributes containing clustered regions from
which outliers can deviate.
Furthermore, as we aim at detecting outliers that deviate
from clustered objects in their local neighborhood we check
uniformly distribution according to this neighborhood N(o,S)
and not w.r.t. the entire subspace. Based on this, our out-
lier ranking is computed for each object w.r.t. its subspace
neighborhood, instead of the whole DB. Hence, the choice of
relevant subspaces occurs strictly on the basis of the object
locality. This is in contrast to subspace search approaches
[25], which provide global subspace estimations supporting
clustering with interesting projections. Compared to such
approaches, the main advantage of our subspace selection is
in the selection of locally significant subspaces for each object
taking local deviation into account.
To illustrate the effects of relevant subspace selection, we
depict a subspace lattice with all possible subspaces of a 4d
data space in Figure 2. Starting considering 1d projections
first, typically these low dimensional projects are uniformly
dense, and thus, they do not affect our outlier ranking. In
the relevant subspace projections highlighted in bold, we find
the detect outliers deviating from their local neighborhood.
As depicted we prune the higher-dimensional subspaces. For
example, adding dimension 3 to subspace {1,2} results in
such an irrelevant subspace. Incrementally using the statistical
test for each dimension we detect the irrelevant dimensions
and stop further processing of higher dimensional subspaces.
Formally, we define relevant subspaces RS(o) in Definition 3
to be the set of subspaces that are significantly non-uniformly
distributed.
Definition 3: Set of Relevant Subspaces
The set of relevant subspaces contains subspaces that are
significantly non-uniformly distributed:
RS(o) = {S ∈ P(D) | S passes H1}
Only these subspaces are considered for our outlier ranking
(cf. Def. 1).
Computing the outlierness of an object in its relevant sub-
spaces RS(o) yields a high contrast between the object and its
local neighborhood. Although we have excluded the irrelevant
subspaces a major challenge remains. Subspaces in RS(o)
have arbitrary dimensionality and show highly varying density
values as motivated in Challenge 1.
C. Adaptive Outlierness in Subspaces
For a meaningful outlier ranking based on outlierness in
multiple subspace projections the definition of score(o,S) has
to provide an adaptive outlierness measure as the overall rank-
ing combines object properties out of very different subspaces
S ∈ RS(o). We propose such an adaptive outlierness measure
by defining an adaptive density and a local deviation for each
object.
1) Adaptive object density: As formalized in Challenge 1,
measuring density in multiple subspaces leads to a challenging
task, namely the strong dependence of densities on the number
of attributes in the considered subspaces. For two subspaces
S,T ⊆ D with T ⊂ S a simple counting of objects in a
fixed neighborhood yields den(o,S) ≤ den(o,T). The main
problem for density-based mining of different subspaces is
the fixed neighborhood [3], [19], [23]. As distances between
objects grow with increasing number of attributes, a fixed
neighborhood N(o,S) = {p | distS(o,p) ≤ ε} becomes
empty. All objects tend to have higher distance than the fixed ε
parameter. To tackle this general problem of density estimation
in arbitrary subspaces, we propose an adaptive density using
a variable neighborhood. By increasing the neighborhood dis-
tance ε with increasing number of attributes, our density mea-
sure can automatically adapt to the expected data distribution.
Thus, different subspaces become comparable and outlierness
based on density estimation can automatically adapt to the
number of attributes.
In general, we propose an adaptive neighborhood AN(o,S)
based on a variable ε(|S|) range.
AN(o,S) = {p | distS(o,p) ≤ ε(|S|)}
The general idea is to derive the variable range out of a
common observation in subspace projections. While increasing
the number of attributes in a subspace projection the volume
of a fixed neighborhoods decreases significantly compared to
the overall volume of the subspace. For example, consider
the volume of the neighborhood covering the whole data
range 0...1 of one attribute with ε = 0.5. If one keeps the
neighborhood range fixed, the volume in subspace S is given
by
vol(N(o,S)) =
π|S|/2
Γ(|S|/2 + 1)· 0.5|S|
with the gamma function Γ(n + 1) = n · Γ(n),Γ(1) =
1,Γ(1/2) =
number of attributes.
√π. The volume decreases with increasing
vol(N(o,S)) ? vol(N(o,T)) for |T| < |S|
.
Thus, the expectation of detecting objects in such neigh-
borhoods is decreasing as well, resulting in very low density
estimations. Our variable neighborhood range adapts to this
phenomenon. While for 1-dimensional subspaces neighbor-
hood rages are typically set to lower values ε ≤ 0.5, in
higher-dimensional subspaces the range should be increased
with the number of dimensions. By increasing the range we
ensure that the expected number of objects remains constant.
Thus, we provide a comparable density estimation in arbitrary

Page 7

subspaces. In our prior work, we have shown that such an
adaptive neighborhood can be of general benefit for other
mining paradigms as well, e.g. for density-based subspace
clustering [19].
2) Instantiation of adaptive density: In the following we
instantiate the basic idea of adaptive neighborhoods to a
specific density estimation technique. As a flexible outlier
model, OUTRES could be used with any density measure such
as the simple counting of objects in the objects neighborhood
(cf. Section III-A). However, we base our density measure
on more enhanced and well established density estimation
techniques [26]. As the overall density distribution of the data
is not known in advance, density den(o,S) of an object o can
be estimated by using kernel density estimators.
Each object o contributes to the overall density by a local
impact defined by a kernel function K(x) with x =distS(o,p)
being the scaled distance of any other object p to the object
o. The bandwidth parameter h is used to scale the influence
of each object to a maximal distance of h. The overall density
for an object o is then simply the sum of kernel function
over all objects in the database. As kernel function we use
the Epanechnikov Kernel
h
Ke(x) = (1 − x2), x < 1
, providing optimal density estimation according to the mean
integrated squared error [26]. Concludingly, den(o,S) is cal-
culated by the formula:
?
Since objects being farther away than h from a certain object
do not contribute to its density, we obtain a local density on
which our outlier detection is based.
In contrast to the simple counting of objects in the neighbor-
hood given by |AN(o,S)|, kernel density estimation has major
benefits due to the weighted influence of each object. The
sum of Epanechnikov Kernels provides a theoretically sound
density definition. However, for density estimation in arbitrary
subspace projections the fixed bandwidth h shows similar
drawbacks to the fixed ε range. For comparable outlierness
over arbitrary subspaces, we propose to adapt the density by
a variable kernel bandwidth ε(|S|). As the true underlying
density distribution is unknown, we only use the dimension-
ality of the space to derive the bandwidth for adaptive density
estimation. For a fixed space with dimensionality d optimal
bandwidth hoptimal(d) is given by the following formula:
?
π
den(o,S) =
1
|DB|
p∈DB
Ke
?distS(o,p)
h
?
hoptimal(d) =
8 · Γ(d
2+ 1)
d
2
· (d + 4) · (2√π)d
?
· n
−1
d+4
where n = |DB| is the database size and the gamma function
as in the computation of the neighborhood volume [26].
As motivated in the previous paragraph, optimal bandwidth
is computed based on the expected number of objects in a
neighborhood. The formula for optimal bandwidth can simply
be seen as the optimal radius of an ε-sphere such that one yield
statistically optimal density estimation results. For density
estimation in subspaces this means that one has to choose
a bandwidth for each individual subspace. Assuming that
n is fixed in a static database, we observe hoptimal(d) to
be a monotonically increasing function. So intuitively, for
increasing dimensionality the influence (bandwidth) of each
object is increased as well in order to maintain optimal
density estimates, while the data space is becoming sparse.
For comparable outlierness we use the optimal bandwidth
to adapt density estimation in arbitrary subspaces. By the
user parameter ε we allow the user to quantify a notion of
locality and adjust this value for arbitrary subspaces based on
the optimal bandwidth. Formally, the bandwidth for a given
number of attributes in a subspace S is defined by Definition 4:
Definition 4: Adaptive neighborhood
For a subspaces dimensionality |S|, |S| ≥ 2, the adaptive
neighborhood ε(|S|) is defined by
ε(|S|) = ε ·hoptimal(|S|)
hoptimal(2)
Thus, we simply scale the given starting bandwidth ε from
2d space up to full data space and use these value for density
estimation. In contrast to the fixed bandwidth in kernel density
estimation we use our adaptive neighborhood as variable
bandwidth for each individual subspace. Consequently, our
automatic bandwidth adaption ensures comparable density
estimates for arbitrary dimensional subspaces.
3) Local object deviation: For our outlier ranking based
on deviations of density we first compute the density for
each object and compare it with the local (average) density
in a relevant subspace. By that, our approach is able to
detect objects highly deviating from the residual data in a
relevant subspace, i.e., objects having exceptionally low den-
sities. While our adaptive density ensures comparability over
multiple subspaces, our local deviation ensures meaningful
outlierness values inside one subspace. Hence, in addition to
the adaptive density, we ensure to highlight an outlier with
very low density compared to its local neighborhood in the
considered subspace.
Having such a comparable density estimation, an outlier can
be detected as an object showing significantly low density. As
we aim at a local outlierness we measure deviation based on
an adaptive threshold. As first filter step we select only objects
with significantly low density
den(o,S) < μ − 2 · σ
compared to μ and σ as the mean and standard deviation of
den(o,S) in the neighborhood of object o. From statistical
observations, only very rare objects deviate more than two
standard deviations from the mean value (cf. Chebyshev’s
inequality [8]). As statistical probability for such objects is
low (e.g. for normal distributed data it is less than 2.1%),
their outlierness has to be high. Using mean μ and standard
deviation σ of the estimated (local) density we ensure to be

Page 8

adaptive to varying density. Object deviation is then given by
the following definition.
Definition 5: Object deviation
The deviation of an object o with respect to mean and standard
deviation of the estimated density:
dev(o,S) =μ − den(o,S)
2 · σ
An object shows high deviation if its density compared
to the average density μ in its neighborhood AN(o,S) is
significantly low.
4) Adaptive Outlierness: Overall the outlierness of an ob-
ject o has to fulfill two major requirements. First, it has to be
adaptive to arbitrary dimensional subspaces. Thus, based on
our adaptive object density we propose an adaptive outlierness
which is comparable for different subspaces (cf. Sec. III-C2).
Second, our adaptive outlierness has to cope with object
deviation considering statistically deviation from the mean
value (cf. Sec. III-C3). Incorporating both aspects in our
adaptive outlierness measure we define score(o,S) as follows:
Definition 6: Adaptive Outlierness
The outlierness of an object o in subspace S is derived by its
density and its deviation in this subspace:
?
1
Our novel outlierness incorporates both aspects derived by
density and the deviation of each object: low density and
high deviation are both indicates for high outlierness. Highly
deviating objects show up by dev(o,S) ≥ 1 as density is
significantly low compared to mean and standard deviation.
Overall we cope with the different behaviors of objects
in different subspaces: Scattered irrelevant subspaces are ex-
cluded by our relevance testing (cf. Sec. III-B). Objects in
a dense subspace S result only in high density and almost
no deviation such that we set score(o,S) = 1 they do not
affect the ranking value (cf. Def. 1). Only if objects show up
with low density or high deviation in a relevant subspace they
contribute to the overall ranking value (cf. Def. 6).
score(o,S) =
den(o,S)
dev(o,S)
, if dev(o,S) ≥ 1
, else.
D. Computation of OUTRES
For the overall computation of our outlier ranking, first of
all one has to select the relevant subspaces for each object. In a
naive solution, one would test for each object o ∈ DB its local
neighborhood AN(o,S) in arbitrary subspaces S ∈ P(D).
As the number of possible subspaces increases exponentially
with the number of given attributes in D this selection is
obviously not practically feasible. For each object, one would
compute RS(o) out of 2|D|possible subspaces by using our
relevance test. Furthermore, we require for all objects a density
computation which yields a quadratic complexity w.r.t. the
number of objects. Overall the complexity of outlier ranking
based on relevant subspaces would be O(|DB|·(2|D|·|DB|)).
Thus, for an efficient processing we propose an approx-
imative selection of relevant subspaces based on a pruning
heuristic. We process subspaces bottom-up and prune based
on the observation that having reached a sparse subspace
with uniformly distributed data we may stop processing, as
data is scattered even more in higher dimensional projections.
As most important property, this pruning ensures to exclude
all irrelevant subspace. This provides a high contrast for our
outlier ranking as all considered subspaces are non-uniformly
distributed. As an approximative selection we cannot ensure
to include all non-uniformly distributed subspaces, but as
highlighted also by our experiments we achieve high quality
outlier ranking with this simple pruning heuristic. We even
observe some redundant subspaces that actually are selected
for RS(o), but do not contribute to the overall outlier ranking.
Thus, further enhancements for algorithmic solutions of our
novel outlier ranking model seem to be promising tasks for
future research. Especially, pruning of dense regions that
do not show any hidden outliers might lead to even better
runtimes. In addition to our pruning of irrelevant subspaces
one might add a second filter to prune some of the relevant
subspaces that might contribute only scores equal to 1 for the
outlierness measure. Such an enhanced filtering could safely
exclude further parts of the search space without any quality
losses.
Algorithm 1 OUTRES(o, S)
FOREACH i ∈ D \ S
S?= S ∪ {i};
IF S?is relevant
den(o,S?) =
dev(o,S?) =μ−den(o,S?)
IF dev(o,S?) ≥ 1
r(o) = r(o) ·den(o,S?)
OUTRES(o, S?);
ELSE
// break recursion for higher dimensional subspaces
// relevance test (cf. Def. 2)
p∈AN(o,S?)Ke(distS?(o,p)
;
// high deviation (cf. Def. 6)
dev(o,S?);
// recursively next subspace
1
|DB|
?
2·σ
ε(|S?|)
);
// aggregation of scoring
For the OUTRES algorithm, we test the relevance of sub-
spaces in a bottom-up processing. We base on the observation
that objects are dense in low dimensional spaces, while for
higher dimensional spaces they diverge until they form an
uniformly distributed scattered space. In Figure 2 we show
a box-plot for the varying distribution of density in various
dimensions as a toy example. Our bottom-up processing starts
with low dimensional subspaces and tests step-by-step each
subspace. The computation of the ranking r(o) based on the
relevant subspaces of the object o is given in Algorithm 1. We
start for each object o the recursive processing OUTRES(o,{}),
recursion stops if an irrelevant subspace has been detected
for object o. This ensures to exclude all irrelevant subspace
projections, which would hinder the distinction of outliers in
our ranking. In the worst case (if for all objects all subspaces
are relevant) complexity remains O(2|D|· |DB|2), but in
practical cases we yield efficient processing as we show in
Section IV in addition to the high quality outlier ranking
results.

Page 9

IV. EXPERIMENTS
We demonstrate the quality of our OUTRES approach on
both synthetic and real world data. We compare OUTRES to
the well established LOF [6] and its recent extensions ABOF
[15] as full space approaches. Furthermore, we compare
against OutRank [20] and SOD [13] as the most recent outlier
rankings based on subspace projections.
For comparability, we implemented all algorithms in our
open-source framework [21]. By extending the popular WEKA
framework we base our work on a widely used data input
format for repeatable and expandable experiments. We used
original implementations provided by the authors and best-
effort re-implementations based on the original papers. We en-
sure comparable evaluations and repeatability of experiments,
as we deploy all implemented algorithms on our website1.
With our SOREX system [22], we ensure that all of our results
will be reproducible, publicly available, and thus, might be
used for comparison in future publications.
For fair comparison we base on objective quality measures.
We believe that our evaluation setup provides a better quality
assessment than showing only some examples of the detected
outlier. In contrast to such a commonly used subjective eval-
uation, we highlight the achieved quality enhancement by
three different quality measures. We measure true positive
(TPR) and false positive (FPR) rates visualized in the well
established ROC plot. Both of these measures are useful to
derive if a ranking detects a high ratio of correct detected
outliers (TPR) while providing only few non-outlier as de-
tected outliers (FPR). However, they only take the ratio of
detected outliers and non-outliers into account ignoring more
or less the positioning of the objects in the ranking. Thus,
we additionally evaluate the results with a ranking coefficient
based on Spearmans Ranking Coefficient [27]. In contrast
to the ROC plot, ranking coefficients take also the ranking
positions of detected outliers into account. This leads to a
more fine grained quality measure.
To illustrate the quality of the rankings we use the quality
measures for the top-k ranked objects (cf. Definition 7). The
TPR measure is simply the fraction of found true outliers in
the first k objects
found true outliers(R,k) =
{or1...ork} ∩ DBhidden outliers
compared to the set of hidden outliers DBhidden outliersin
the database DB. Analogue, FPR is the fraction of found
non-outliers in the first k objects
found false outliers(R,k) =
{or1...ork} \ found true outliers(R,k)
compared to the overall set of non-outlier objects in the
database.
1http://dme.rwth-aachen.de/OpenSubspace/SOREX
Definition 7: TPR and FPR measures
The true positive rate for the first k objects of a ranking R =
{or1...orn} is defined as:
TPR(R,k) =|found true outliers(R,k)|
|DBhidden outliers|
The false positive rate is defined as:
FPR(R,k) =|found false outliers(R,k)|
|DBhidden non-outliers|
More detailed measures can be derived by ranking coefficients
[27]. Spearmans Ranking Coefficient SRC(R1,R2) computes
the correlation of two given rankings R1 and R2. We use
SRC to measure the quality for one ranking by comparing
it with the optimal ranking Rbest, ranking all outliers first.
Furthermore, we normalize with the ranking coefficient for the
worst ranking Rworsthaving all outliers in the last positions.
We define outlier ranking coefficient ORC(R,k) for the first
k objects in ranking R as given in Definition 8.
Definition 8: Ranking coefficient measure
The outlier ranking coefficient for the first k objects of a
ranking R = {or1...orn} is defined as:
ORC(R,k) =SRC({or1...ork},Rbest)
SRC(Rworst,Rbest)
For these measures the optimal ranking results in
TPR(Roptimal,k) = 1 ∧ FPR(Roptimal,k) = 0 and
ORC(Roptimal,k) = 1 for k = |DB|outliers
For non-optimal rankings TPR = 1 is reached for larger k
with FPR ? 0, while the ORC measure does not reach
the maximal value of 1 at all for non-optimal rankings. Thus,
the ORC measure is more appropriate for evaluation of outlier
rankings. By taking the actual positing of objects into account,
ORC is able to distinguish between two rankings having
found the same amount of outliers in the first k positions.
In such a case, TPR and FPR show same results as they
only consider the object ratio and cannot distinguish between
these two rankings. Taking also positioning information into
account ORC shows more fine grained differences in rank-
ings. Especially, one can compare ranking quality by taking the
overall ORC(R,|DB|) for comparison. Thus, after showing
all three measures in the first experiment we use only the
ranking coefficient measure for comparison in the following
experiments.
A. Synthetic Data
For scalability experiments, we generate synthetic data
following a method proposed in [11], [3] to generate density-
based clusters in arbitrary subspaces. In addition, our generator
adds outliers deviating from one of these subspace clusters. As
there are no global patterns hidden in data, the hidden outliers
do not appear in the scattered full space. In our first exper-
iment, we evaluate the quality of the competing approaches

Page 10

on a synthetic data set with 4765 objects represented by four
subspace clusters each using 4 out the 16 given attributes and
additionally 61 hidden outliers deviating from these clusters.
Figure 3(a) illustrates the quality with respect to ROC plot.
We observe that all approaches show high increase in true
positive rates of detected outliers with only very few false
positive. However, all hidden outliers (TPR = 1) are found
after thousands of considered objects, indicated by FPR ? 0.
Our novel OUTRES shows best performance compared to
LOF, ABOF, SOD and OutRank, as it archives to detect more
hidden outliers within the first ranked objects showing both
higher TPR and lower FPR than the competitors. Com-
paring ROC plot and ranking coefficient in Figure 3 for the
same experiment, we observe that OUTRES outperforms all
competing approaches in both quality measures independent
of the number of ranked objects. For ranking coefficient it
always shows highest correlation with the optimal ranking.
As ranking coefficient provides more information about the
positioning of the objects than the ROC plot, we use only this
measure in the following experiments.
0
0,2
0,4
0,6
0,8
1
0 0,2 0,40,60,81
True Positive Rate
False Positive Rate
(a) ROC plot
OUTRES
OutRank
LOF
ABOF
SOD
0
0,2
0,4
0,6
0,8
1
0 10002000 3000 4000
ranking coefficient
ranked objects
OUTRES
OutRank
LOF
ABOF
SOD
(b) ranking coefficient
Fig. 3.Ranking quality on synthetic data
In our second experiment, we evaluate the scalability of
outlier rankings with respect to the number of given attributes
in the database. As outlier ranking in subspace projections
aims to detect outliers hidden in any subset of the given at-
tributes, scalabiltiy w.r.t. number of given attributes is crucial.
We varied the number of attributes from 10 up to 50, while
keeping number of hidden subspace clusters, hidden outliers
and database size constant, as in the previous experiment.
Figure 4(a) shows the almost constantly high quality of our
approach. As we add more an more attributes, hidden outliers
disappear in the overall scattered full space. However, as
OUTRES investigates only relevant subspace projections it
scales w.r.t. number of given attributes. It outperforms all
competing approaches in terms of quality.
1
0,8
nt
oefficien
king?co
ran
0,6
0,4
OUTRES
OutRank OutRank
LOF
ABOF
ABOF
SOD
0,2
00
1020 30 40 50
data?space?dimensionality
(a) ranking quality
1
10
100
1000
10000
100000
1000000
10 2030 4050
runtime [sec.]
data space dimensionality
ABOF
OUTRES
SOD
OutRank
LOF
(b) runtime
Fig. 4. Scalability w.r.t. number of attributes
In Figure 4(b) we compare the runtimes. In conjunction
with the previous quality plot we observe that OUTRES
achieves to perform both efficient outlier ranking and a high
outlier ranking quality. Although OUTRES has to search for
outliers in arbitrary subsets of the attributes, the proposed
pruning heuristic by excluding uniformly distributed attributes
shows both high quality results but also efficient computation.
We skip further scalability experiments, as experiments have
shown that database size has less impact on both quality
and runtime. However, as an important issue we discuss
parametrization of our approach in the following experiments.
B. Parametrization
For the two main parameters α and ε we show the robust-
ness of the ranking quality of OUTRES. On the synthetic data
set from previous experiment, we varied the neighborhood
parameter ε from 5 to 45 (data ranges from 0 to 100).

Page 11

As depicted in Figure 5(a), OUTRES shows a quite robust
ranking quality only slightly decreasing for high ε values.
By increasing the neighborhood around each object density
is increasing for all objects. Especially for outliers, density is
becoming similar to clustered objects. Overall we achieve a
robust approach w.r.t. ε due to our automatic adaption of the
neighborhood range for the arbitrary subspace projections con-
sidered in OUTRES (cf. Def. 4). As default setting of ε in our
experiments we use ε = 15 showing best results. In general,
with our adaptive neighborhood one can set the usual low ε
neighborhood ranges in low dimensional subspace. These are
increased automatically for higher-dimensional subspaces and
provide a high quality density estimation for arbitrary subspace
projections.
0,5
0,4
0,6
0,7
0,8
0,9
1
coefficient
ranking?
OUTRES
0
0,1
0,2
0,3
5 10 1520 25 3035 4045
? ? parameter
(a) Variation of ε
0,5
0,4
0,6
0,7
0,8
0,9
1
coefficient
ranking?
OUTRES
0
0,1
0,2
0,3
0 0,10,20,30,4
? ? parameter
(b) Variation of α
Fig. 5. Robustness of OUTRES w.r.t parameters
For the second parameter α we observe similar effects. As
depicted in Figure 5(b) we observe best ranking quality for
low α settings. For higher α settings, OUTRES accepts more
and more uniform distributed subspaces as relevant subspaces
for outlier ranking. As one cannot distinguish between outliers
and regular objects in these scattered subspaces, the overall
ranking quality decreases. Keeping a low α setting (default
α = 0.01), thus, ensures to measure outlierness of objects
only in relevant subspace projections.
C. Real World Data
We analyzed the quality of outlier ranking on three real
world data sets (Ionosphere, Breast Cancer and Pendigits) from
the UCI repository [7]. All of these data sets provide scattered
full spaces, while subsets of the given attributes can be used to
distinguish between the hidden patterns and outlying objects.
For example, in the Pendigits data set objects are described by
(x,y) positions concatenated in a digit trajectory. Clearly not all
of the pen positions are important to detect outlying objects.
Some digits deviate significantly in the first position (first two
given attributes) from the residual objects starting typically at
similar positions (upper left area). Similarly, also in the other
data sets outliers can be distinguished from regular objects
using subspace projections of the database. For our evaluation
measures, we used one of the class labels reduced to 10% of its
size as ground truth for hidden outliers. In contrast to adding
artificial outliers into the database, such a reduction of the
orginal data distribution seems more natural. The remaining
objects of the reduced class show high deviation from other
classes and have low density due to the eliminated parts of
their own class. In our experiments, we show that outlier
ranking approaches successfully detects these very rare hidden
observations in subspaces of the given databases.
In Figure 6 we show the ranking coefficients for the real
world databases. For all data sets we observe a high ranking
quality of OUTRES, outperforming competing approaches by
detecting outliers as top ranked objects. Hidden outliers are
clearly distinguished by the selection of relevant subspaces.
For example, in the pendigits data one object representing the
digit four has been first ranked due to its high deviation in
the last positions of its trajectory. Overall, our novel OUTRES
approach provides highest quality results, while the competing
approaches show varying quality over multiple data sets.
V. CONCLUSION
In this work, we proposed a novel outlier ranking for objects
deviating in subspace projections. The OUTRES approach
computes local density deviation by looking at a selection of
relevant subspaces for each object. Relevance of subspaces is
measured by statistical significance tests. Thus, only relevant
subspaces that are not distributed uniformly random are used
for our outlier ranking. For comparable outlierness measures
in different subspaces, we derive an adaptive density mea-
sure which automatically adapts to the considered subspace.
OUTRES computes an overall high quality outlier ranking
by aggregating this adaptive outlierness of objects in relevant
subspaces. Our thorough evaluation on both synthetic and real
world data shows that OUTRES outperforms competing outlier
ranking approaches. Especially, OUTRES achieves to detect
outliers hidden in subspace projections.
ACKNOWLEDGMENT
This work has been supported in part by the UMIC Research
Centre, RWTH Aachen University, Germany.

Page 12

REFERENCES
[1] C. C. Aggarwal and P. S. Yu, “Outlier detection for high dimensional
data,” in SIGMOD, 2001, pp. 37–46.
[2] R. Agrawal, J. Gehrke, D. Gunopulos, and P. Raghavan, “Automatic sub-
space clustering of high dimensional data for data mining applications,”
in SIGMOD, 1998, pp. 94–105.
[3] I. Assent, R. Krieger, E. M¨ uller, and T. Seidl, “DUSC: Dimensionality
unbiased subspace clustering,” in ICDM, 2007, pp. 409–414.
[4] V. Barnett and T. Lewis, Outliers in statistical data. John Wiley, 1994.
[5] K. Beyer, J. Goldstein, R. Ramakrishnan, and U. Shaft, “When is nearest
neighbors meaningful,” in IDBT, 1999, pp. 217–235.
[6] M. Breunig, H.-P. Kriegel, R. Ng, and J. Sander, “LOF: Identifying
density-based local outliers,” in SIGMOD, 2000, pp. 93–104.
[7] A. Frank and A. Asuncion, “UCI machine learning repository,” 2010.
[Online]. Available: http://archive.ics.uci.edu/ml
[8] G. Hardy, J. Littlewood, and G. Polya, Inequalities.
University Press, 1988.
[9] Z. He, X. Xu, and S. Deng, “Discovering cluster-based local outliers,”
Pattern Recognition Letters, vol. 24, no. 9-10, pp. 1641–1650, 2003.
[10] I. Joliffe, Principal Component Analysis.
[11] K. Kailing, H.-P. Kriegel, and P. Kr¨ oger, “Density-connected subspace
clustering for high-dimensional data,” in SDM, 2004, pp. 246–257.
[12] E. Knorr, R. Ng, and V. Tucakov, “Distance-based outliers: algorithms
and applications,” VLDB Journal, vol. 8, no. 3, pp. 237–253, 2000.
[13] H.-P. Kriegel, P. Kr¨ oger, E. Schubert, and A. Zimek, “Outlier detection
in axis-parallel subspaces of high dimensional data,” in PAKDD, 2009,
pp. 831–838.
[14] H.-P. Kriegel, P. Kr¨ oger, and A. Zimek, “Clustering high-dimensional
data: A survey on subspace clustering, pattern-based clustering, and
correlation clustering,” ACM TKDD, vol. 3, no. 1, pp. 1–58, 2009.
[15] H.-P. Kriegel, M. Schubert, and A. Zimek, “Angle-based outlier detec-
tion in high-dimensional data,” in KDD, 2008, pp. 444–452.
[16] A. Lazarevic and V. Kumar, “Feature bagging for outlier detection,” in
KDD, 2005, pp. 157–166.
[17] F. T. Liu, K. M. Ting, and Z.-H. Zhou, “Isolation forest,” in ICDM,
2008, pp. 413–422.
[18] G. Moise and J. Sander, “Finding non-redundant, statistically significant
regions in high dimensional data: a novel approach to projected and
subspace clustering,” in KDD, 2008, pp. 533–541.
[19] E. M¨ uller, I. Assent, S. G¨ unnemann, R. Krieger, and T. Seidl, “Rel-
evant Subspace Clustering: mining the most interesting non-redundant
concepts in high dimensional data,” in ICDM, 2009, pp. 377–386.
[20] E. M¨ uller, I. Assent, U. Steinhausen, and T. Seidl, “Outrank: Ranking
outliers in high dimensional data,” in DBRank Workshop, 2008, pp. 600–
603.
[21] E. M¨ uller, S. G¨ unnemann, I. Assent, and T. Seidl, “Evaluating clustering
in subspace projections of high dimensional data,” in VLDB, 2009, pp.
1270–1281.
[22] E. M¨ uller, M. Schiffer, P. Gerwert, M. Hannen, T. Jansen, and T. Seidl,
“SOREX: Subspace outlier ranking exploration toolkit,” in ECML
PKDD, 2010, pp. 607–610.
[23] E. M¨ uller, M. Schiffer, and T. Seidl, “Adaptive outlierness for subspace
outlier ranking,” in CIKM, 2010, pp. 1629–1632.
[24] S. Papadimitriou, H. Kitagawa, P. Gibbons, and C. Faloutsos, “LOCI:
Fast outlier detection using the local correlation integral,” in ICDE, 2003,
pp. 315–326.
[25] L. Parsons, E. Haque, and H. Liu, “Subspace clustering for high
dimensional data: a review,” SIGKDD Explor. Newsl., vol. 6, no. 1, pp.
90–105, 2004.
[26] B. Silverman, Density Estimation for Statistics and Data Analysis.
Chapman and Hall, London, 1986.
[27] C. Spearman, “The proof and measurement of association between two
things,” American Journal of Psychology, vol. 15, no. 1, pp. 72–101,
1987.
[28] M. Stephens, “Use of the Kolmogorov-Smirnov, Cramer-von Mises
and related statistics without extensive tables,” Journal of the Royal
Statistical Society. Series B, pp. 115–122, 1970.
Cambridge
Springer, New York, 1986.
0,4
0,3
0,5
0,6
0,7
0,8
0,9
1
nking?coefficient
ran
OUTRES
OutRank
LOF
ABOF
SOD
0
0,1
0,2
0 1000 20003000
ranked?objects
400050006000 7000
(a) Pendigits (digit 4)
0,4
0,3
0,5
0,6
0,7
0,8
0,9
1
nking?coefficient
ran
0
0,1
0,2
0 10002000 3000
ranked?objects
4000500060007000
OUTRES
OutRank
LOF
ABOF
SOD
(b) Pendigits (digit 6)
1
0,8
nt
oefficien
king?co
ran
0,6
0,4
OUTRES
OutRank
LOF
ABOF
SOD
0,2,
00
0 50100
ranked?objects
150200 250
(c) Ionosphere
1
0,8
nt
oefficien
king?co
ran
0,6
0,4
OUTRES
OutRank OutRank
LOF
ABOF ABOF
SOD
0,2
00
0 50 100150
ranked?objects
(d) Breast Cancer
Fig. 6. Ranking quality on real world data