Algorithms for Spatial Outlier Detection

Chang-Tien Lu Dechang Chen Yufeng Kou

Dept. of Computer Science Preventive Medicine and Dept. of Computer Science
Virginia Polytechnic Institute Biometrics Virginia Polytechnic Institute
and State University Uniformed Services University and State University
7054 Haycock Road of the Health Sciences 7054 Haycock Road
Falls Church, VA 22043 Bethesda, MD 20814 Falls Church, VA 22043
ctlu@vt.edu dchen@usuhs.mil ykou@vt.edu

Abstract outlier is a local instability, or an extreme observation with

respect to its neighboring values, even though it may not be
A spatial outlier is a spatially referenced object whose significantly different from the entire population. Detect-
non-spatial attribute values are significantly different from ing spatial outliers is useful in many applications of geo-
the values of its neighborhood. Identification of spatial out- graphic information systems, including transportation, ecol-
liers can lead to the discovery of unexpected, interesting, ogy, public safety, public health, climatology, and location
and useful spatial patterns for further analysis. One draw- based services [12].
back of existing methods is that normal objects tend to be Recent work by Shekhar et al. introduced a method for
falsely detected as spatial outliers when their neighborhood detecting spatial outliers in graph data set [13, 14]. The
contains true spatial outliers. In this paper, we propose method is based on the distribution property of the differ-
a suite of spatial outlier detection algorithms to overcome ence between an attribute value and the average attribute
this disadvantage. We formulate the spatial outlier detec- value of its neighbors. Several spatial outlier detection
tion problem in a general way and design algorithms which methods are also available in the literature of spatial statis-
can accurately detect spatial outliers. In addition, using tics. These methods can be generally grouped into two cat-
a real-world census data set, we demonstrate that our ap- egories, namely graphic approaches and quantitative tests.
proaches can not only avoid detecting false spatial outliers Graphic approaches are based on visualization of spatial
but also find true spatial outliers ignored by existing meth- data which highlights spatial outliers. Example methods
ods. include variogram clouds and pocket plots [3, 10]. Quan-
titative methods provide tests to distinguish spatial outliers
from the remainder of data. Scatterplot [2, 8] and Moran
scatterplot [9] are two representative approaches.
1 Introduction One major drawback of the existing detection ap-
proaches is that their application will lead to some true
Outliers have been informally defined as observations in spatial outliers being ignored and some false spatial out-
a data set which appear to be inconsistent with the remain- liers being identified. To minimize such defect, we propose
der of that set of data [1, 5], or which deviate so much from two iterative algorithms that detect spatial outliers by multi-
other observations so as to arouse suspicions that they were iterations. Each iteration identifies only one outlier and
generated by a different mechanism [4]. The identification modifies the attribute value of this outlier so that this out-
of outliers can lead to the discovery of useful knowledge lier will not impact the subsequent iterations negatively. We
and has a number of practical applications in areas such also propose a non-iterative algorithm which uses the me-
as credit card fraud detection, athlete performance analy- dian as the neighborhood function, thus reducing the neg-
sis, voting irregularity analysis, and severe weather predic- ative impact caused by the presence of neighboring points
tion [6, 11, 16]. In a spatial context, local anomalies are with very high/low attribute values. Using a real-world cen-
of paramount importance. Spatial outliers are spatially ref- sus data, we show that our algorithms can avoid detecting
erenced objects whose non-spatial attribute values are sig- false spatial outliers and can find true spatial outliers ig-
nificantly different from those of other spatially referenced nored by existing methods when the expected number of
objects in their spatial neighborhoods. Informally, a spatial spatial outliers is limited.

Proceedings of the Third IEEE International Conference on Data Mining (ICDM’03)

0-7695-1978-4/03 $ 17.00 © 2003 IEEE
 1


2 Problem Formulation = k x∈N Nk (xi ) f (x) and the comparison function
hi = h(xi ) = fg(x
(xi )
i)
.
Given a set of spatial points X = {x1 , x2 , . . . , xn } in
a space with dimension p ≥ 1, an attribute function f is 2. Let hq or h−1q denote the maximum of h1 , h2 ,. . ., hn ,
defined as a mapping from X to R (the set of real num-
h−1
1 , h −1
2 , . . ., h−1
n . For a given threshold θ, if hq or
ber). Attribute function f (xi ) represents the attribute value −1
hq ≥ θ, treat xq as an S-outlier.
of spatial point xi . For a given point xi , let N Nk (xi ) de-
note the k nearest neighbors of point xi , where k = k(xi ) 3. Update f (xq ) to be g(xq ). For each spatial point xi
depends on the value of xi for i = 1, 2, . . . , n. A neighbor- whose N Nk (xi ) contains xq , update g(xi ) and hi .
hood function g is defined as a map from X to R such that
for each xi , g(xi ) returns a summary statistic of attribute 4. Repeat steps 2 and 3 until either the threshold condi-
values of all the spatial points inside N Nk (xi ). For exam- tion is not met or the total number of S-outliers equals
ple, g(xi ) can be the average attribute value of the k nearest m.
neighbors of xi . To detect spatial outliers, we compare the
attribute value of each point xi with those attribute values of In Algorithm 2 below, the neighborhood function g is
its neighbors N Nk (xi ). Such comparison is done through the same as in Algorithm 1. But the comparison func-
a comparison function h, which is a function of f and g. tion h(x) is chosen to be the difference f (x) − g(x). Ap-
There are many choices for the form of h. For example, h plying such an h to the n spatial points leads to the se-
can be the difference f − g or the ratio f /g. Let yi = h(xi ) quence {h1 , h2 , . . . , hn }. A spatial point xi is treated as
for i = 1, 2, . . . , n. Given the attribute function f , function a candidate of S-outlier if its corresponding value hi is
k, neighborhood function g, and comparison function h, a extreme among the data set {h1 , h2 , . . . , hn }. Let µ and
point xi is a spatial outlier or simply S-outlier if yi is an σ denote the sample mean and sample standard deviation
extreme value of the set {y1 , y2 , . . . , yn }. We note that the of {h1 , h2 , . . . , hn }. The standardized value for each hi
definition depends on the choices of functions k, g and h. is yi = hiσ−µ , and so the standardized data set becomes
The definition given above is quite general. As a matter {y1 , y2 , . . . , yn }. Now it is clear that xi is extreme in the
of fact, outliers involved in various existing spatial outlier original data set iff hi is extreme in the standardized data
detection techniques are special cases of S-outliers [15]. set. Correspondingly, xi is a possible S-outlier if |yi | is
These include outliers detected by z algorithm [13], Scat- large enough (again detected by θ).
terplot [2, 8], Moran scatterplot [9] and pocket plots [3, 10]. Algorithm 2 (Iterative z Algorithm)

1. For each spatial point xi , compute the k nearest neigh-

3 Proposed Algorithms

N Nk (xi ), the neighborhood function g(xi )
bor set
= k1 x∈N Nk (xi ) f (x), and the comparison function
We state our algorithms to detect S-outliers. For sim-
hi = h(xi ) = f (xi ) − g(xi ).
plicity, the description assumes all k(xi ) are equal to a fixed
number k. The algorithms can be easily generalized by re- 2. Let µ and σ denote the sample mean and sample stan-
placing fixed k with dynamic k(xi ). As seen above, outlier dard deviation of the data set {h1 , h2 , . . . , hn }. Stan-
detection algorithms depend on the choices of the neighbor- dardize the data set and compute the absolute value
hood function g and comparison function h. Selection of g yi = | hiσ−µ | for i = 1, 2, . . . , n. Let yq denote the
and h determines the performance of each algorithm. In Al- maximum of y1 , y2 , . . . , yn . For a given threshold θ, if
gorithm 1 below, the neighborhood function g evaluated at yq ≥ θ, treat xq as an S-outlier.
a spatial point x is taken to be the average attribute value
of all the k nearest neighbors of x. Comparison function 3. Update f (xq ) to be g(xq ). For each spatial point xi
h(x) is taken to be the ratio of f (x) to g(x). Very large whose N Nk (xi ) contains xq , update g(xi ) and hi .
or very small value h(x) (detected by the threshold θ) is an
indication that x might be an S-outlier. Algorithm 1 is also 4. Recalculate µ and σ of the data set {h1 , h2 , . . . , hn }.
termed as an iterative r(ratio) algorithm, since iterations are For i = 1, 2, . . . , n, update yi = | hiσ−µ |.
coupled with the ratios.
Algorithm 1 ( Iterative r Algorithm) 5. Repeat steps 2, 3, and 4 until either the threshold con-
dition is not met or the total number of S-outliers
1. Given a spatial data set X = {x1 , x2 , . . . , xn }, an at- equals m.
tribute function f , a number k of nearest neighbors,
and an expected number m of spatial outliers. For The simplest choice of θ in Algorithm 1 is θ = 1. It can
each spatial point xi , compute the k nearest neigh- be larger than 1 depending on different scenarios. A com-
bor set N Nk (xi ), the neighborhood function g(xi ) mon value of θ in Algorithm 2 may be taken to be 2 or 3.

Proceedings of the Third IEEE International Conference on Data Mining (ICDM’03)

0-7695-1978-4/03 $ 17.00 © 2003 IEEE
 This is based on the result that in Algorithm 2 the compar-
ison function h is normally distributed if the attribute func- 200

tion f is normally distributed [15]. There is no clear guide-

line on which of the two algorithms (iterative r and z algo- 150 S1

Attribute Value
rithms) should be selected in applications. If the attribute 100

function f can take negative values, then it is obvious that S2

the iterative r algorithm should not be used due to the cur- 50

rent description of the algorithm. If the attribute function f 0

120
S3

is non-negative, the selection depends on the properties of 100 E1

E2
120
140
80
the practical applications. In general, we recommend that 60
60
80
100

40 40
both algorithms be used. Y coordinate 20 0
20

X coordinate
In both Algorithms 1 and 2, once an S-outlier is detected,
some corrections are made immediately. These include re-
placing the attribute value of the outlier by the average at- Figure 1. A spatial data set. Objects are lo-
tribute value of its neighbors and some subsequent updating cated in the X − Y plane. The height of each
computation. The effect of these corrections is to avoid nor- vertical line segment represents the attribute
mal points close to the true outliers to be claimed as possible value of each object.
outliers. There is a direct method to reduce the risk of over-
Methods
stating the number of outliers without replacing the attribute
Rank Scatter- Moran z Iterative Iterative Median
value of the detected outlier, as describe in Algorithm 3. plot Scatterplot Alg. z Alg. r Alg. Alg.
The method in Algorithm 3 defines the neighborhood func- 1 E1 S1 S1 S1 S1 S1
tion differently. Instead of the average attribute value, g(xi ) 2 E2 E1 E1 S2 S2 S2
is chosen to be the median of the attribute values of the 3 S2 E2 E2 S3 S3 S3
points in N Nk (xi ). The motivation of using median is the
fact that median is a robust estimator of the “center” of a
Table 1. The top three spatial outliers detected
sample.
by Scatterplot, Moran scatterplot, z, iterative
Algorithm 3 (Median Algorithm)
z, iterative r, and median algorithms.
1. For each spatial point xi , compute the k nearest neigh-
bor set N Nk (xi ), the neighborhood function g(xi ) =
median of the data set {f (x) : x ∈ N Nk (xi )}, and the
comparison function hi = h(xi ) = f (xi ) − g(xi ). 4 Experiments
2. Let µ and σ denote the sample mean and sample stan- We empirically compared the detection performance of
dard deviation of the data set {h1 , h2 , . . . , hn }. Stan- our proposed methods with the z algorithm through min-
dardize the data set and compute the absolute values ing a real-life census data set. The experiment results in-
yi = | hiσ−µ | for i = 1, 2, . . . , n. dicate that our algorithms can successfully identify spatial
3. For a given positive integer m, let i1 , i2 , . . . , im be the outliers ignored by the z algorithm and can avoid detect-
m indices such that their y values in {y1 , y2 , . . . , yn } ing false spatial outliers. In this experiment, we tested var-
represent the m largest. Then the m S-outliers are ious attributes from census data compiled by U.S. Census
xi1 , xi2 , . . . , xim . Bureau [17]. The attributes tested include population, pop-
ulation density, percent of white persons, percent of black
A quick illustration of Algorithms 1, 2, and 3 is to apply or African American persons, percent of American Indian
them to the data in Figure 1. Table 1 shows the results using persons, percent of Asian persons, and percent of female
the three algorithms with parameters k = m = 3, com- persons. We first ran the four algorithms (z, iterative r, it-
pared with the existing approaches. As can be seen, all the erative z, and median algorithms) to detect which counties
three proposed algorithms accurately detect S1, S2, and S3 have abnormal population. There are 3192 counties in the
as spatial outliers, but z algorithm, Scatterplot, and Moran USA. We show the top 10 counties which are most likely to
Scatterplot, falsely identify E1 and E2 as spatial outliers. be the spatial outliers.
In this table, the rank of the outliers is defined in an obvi- Table 2 provides the experimental results for all four spa-
ous way. For example, in iterative r and z algorithms, the tial outlier detection algorithms. For the top 10 spatial out-
rank is the order of iterations, while in both z and Median lier detected by z, iterative z, and median algorithms, most
algorithms, the rank is determined by the y value.
of them are the same with slightly different order. In fact,
there are eight spatial outliers in common detected by the

Proceedings of the Third IEEE International Conference on Data Mining (ICDM’03)

0-7695-1978-4/03 $ 17.00 © 2003 IEEE
 Methods
Rank z Alg. Iterative z Alg. Median Alg. Iterative r Alg.
1 Los Angeles,CA,9637494.0 Los Angeles,CA,9637494.0 Los Angeles,CA,9637494.0 Kenedy,TX,413.0
2 Cook,IL,5350269.0 Cook,IL,5350269.0 Cook,IL,5350269.0 Loving,TX,70.0
3 Harris,TX,3460589.0 Harris,TX,3460589.0 Harris,TX,3460589.0 Treasure,MT,802.0
4 Maricopa,AZ,3194798.0 Maricopa,AZ,3194798.0 Maricopa,AZ,3194798.0 Lincoln,NV,4198.0
5 Ventura,CA,770630.0 Dallas,TX,2245398.0 Miami-Dade,FL,2289683.0 Falls Church city,VA,10612.0
6 Dallas,TX,2245398.0 Miami-Dade,FL,2289683.0 Dallas,TX,2245398.0 La Paz,AZ,19759.0
7 Miami-Dade,FL,2289683.0 Wayne,MI,2045473.0 Wayne,MI,2045473.0 Alpine,CA,1192.0
8 Wayne,MI,2045473.0 Bexar,TX,1417501.0 Clark,NV,1464653.0 Hudspeth,TX, 3318.0
9 Bexar,TX,1417501.0 King,WA,1741785.0 Bexar,TX,1417501.0 Fairfax city,VA,21674.0
10 King,WA,1741785.0 San Diego,CA,2862819.0 Tarrant,TX,1486392.0 Gilpin, CO,4823.0

Table 2. The top ten spatial outliers detected by z, iterative z, iterative r, and Median algorithms.

three algorithms. Further examination shows that the top [3] J. Haslett, R. Brandley, P. Craig, A. Unwin, and G. Wills.
10 counties selected by iterative z and median algorithms Dynamic Graphics for Exploring Spatial Data With Appli-
are true outliers. But Ventura Co. in California was falsely cation to Locating Global and Local Anomalies. The Amer-
detected by the non-iterative z algorithm. This falsely de- ican Statistician, 45:234–242, 1991.
[4] D. Hawkins. Identification of Outliers. Chapman and Hall,
tected outlier was avoided by both the iterative z and median
1980.
algorithms. The last column of the table shows the top ten [5] R. Johnson. Applied Multivariate Statistical Analysis. Pren-
candidate outliers from the iterative r algorithm. Although tice Hall, 1992.
these 10 candidates are true outliers from a practical exami- [6] E. Knorr and R. Ng. Algorithms for Mining Distance-Based
nation, they are very different from those obtained from the Outliers in Large Datasets. In Proc. 24th VLDB Conference,
other three methods. This difference is due to the fact that 1998.
the iterative r algorithm focuses on the ratio between the at- [7] C.-T. Lu, H. Wang, and Y. Kou.
tribute value and the averaged attribute value of neighbors. http://europa.nvc.cs.vt.edu/˜ctlu/Project/MapView/index.htm.
[8] A. Luc. Exploratory Spatial Data Analysis and Geographic
Experimental results from other attributes also show that
Information Systems. In M. Painho, editor, New Tools for
the iterative algorithms and median method are more ac- Spatial Analysis, pages 45–54, 1994.
curate than the non-iterative algorithm in terms of falsely [9] A. Luc. Local Indicators of Spatial Association: LISA. Ge-
detected spatial outliers. For running the algorithms and ographical Analysis, 27(2):93–115, 1995.
generating more results, we refer interested readers to [7], [10] Y. Panatier. Variowin. Software For Spatial Data Analysis in
where we developed one software package which imple- 2D. New York: Springer-Verlag, 1996.
ments all the existing and proposed algorithms. [11] I. Ruts and P. Rousseeuw. Computing Depth Contours of Bi-
variate Point Clouds. In Computational Statistics and Data
Analysis, 23:153–168, 1996.
5 Conclusion [12] S. Shekhar and S. Chawla. A Tour of Spatial Databases.
Prentice Hall, 2002.
[13] S. Shekhar, C.-T. Lu, and P. Zhang. Detecting Graph-Based
In this paper we propose three spatial outlier detection Spatial Outlier: Algorithms and Applications(A Summary
algorithms to analyze spatial data: two algorithms based on of Results). In Proc. of the Seventh ACM-SIGKDD Int’l
iteration and one algorithm based on median. The experi- Conference on Knowledge Discovery and Data Mining, Aug
mental results confirm the effectiveness of our approach in 2001.
reducing the risk of falsely claiming regular spatial points as [14] S. Shekhar, C.-T. Lu, and P. Zhang. Detecting Graph-Based
outliers, which exists in commonly used detection method- Spatial Outlier. Intelligent Data Analysis: An International
ologies. Furthermore, it carries the important bonus of or- Journal, 6(5):451–468, 2002.
[15] S. Shekhar, C.-T. Lu, and P. Zhang. A Unified Approach
dering the spatial outliers with respect to their degree of out-
to Spatial Outliers Detection. GeoInformatica, An Interna-
lierness. tional Journal on Advances of Computer Science for Geo-
graphic Information System, 7(2), June 2003.
References [16] T. Johnson and I. Kwok and R. Ng. Fast Computation
of 2-Dimensional Depth Contours. In Proceedings of the
Fourth International Conference on Knowledge Discovery
[1] V. Barnett and T. Lewis. Outliers in Statistical Data. John and Data Mining, pages 224–228. AAAI Press, 1998.
Wiley, New York, 3rd edition, 1994. [17] U.S. Census Burean, United Stated Department of Com-
[2] R. Haining. Spatial Data Analysis in the Social and Envi- mence. http://www.census.gov/.
ronmental Sciences. Cambridge University Press, 1993.

Proceedings of the Third IEEE International Conference on Data Mining (ICDM’03)

0-7695-1978-4/03 $ 17.00 © 2003 IEEE