Crime Hot-Spots Prediction Using Support Vector Machine

Keivan Kianmehr
Department of Computer Science
University of Calgary
Calgary, Alberta, CANADA
kiamehr@cpsc.ucalgary.ca

Reda Alhajj
Department of Computer Science
University of Calgary
Calgary, Alberta, CANADA
alhajj@cpsc.ucalgary.ca

Abstract

into the same class. For instance, if the unemployment in a

city is high and the population of the city is high then the
crime rate in that city is high. Spatial classification deals
with datasets that contain spatial objects. In spatial classification, attribute values of neighboring objects may also be
relevant for the membership of objects in a certain group.
This neighborhood factor which is called spatial autocorrelation plays a significant role in spatial data analysis and
poses challenges in spatial data classification as well [5].
Traditional data mining methods cannot be effectively applied to spatial data classification since they do not consider
spatial autocorrelations among objects. Therefore, there is
the need to integrate the existing techniques into new approaches to the spatial data classification task.
Location prediction is a special case of spatial data classification in which we are interested to predict the location.
For instance, we may be interested to predict the location
of crime hot spots as important part of public safety. The
output of the analysis can provide useful information to improve the activities for preventing and detecting safety and
security problems. The availability of a location prediction
inquiry system can be helpful for public safety experts.
A wide variety of research has considered the use of data
mining techniques like Neural Network [6], logistic regression [7] and decision tree to extract patterns from spatial
datasets to predict location. However, these previous studies have shown that traditional techniques do a poor job in
predicting location task because they do not consider the
spatial relations between spatial objects. Several techniques
have been proposed to guarantee the spatial dependencies of
objects. PLUMS by Shekhar and Chawla [8], is a method
for supervised spatial data classification based on using map
similarity measure. Spatial Autoregressive Regression technique has also been proposed by spatial statisticians [9].
This paper presents one-class support vector machines
(SVMs) based approach [10, 11] to analyze and explore
crime datasets. However, the framework we are providing
here, can be applied to different domains. In our selected
crime datasets, the location is described by Euclidian co-

Location prediction is a special case of spatial data mining classification. For instance, in the public safety domain, it may be interesting to predict location(s) of crime
hot spots. In this study, we present Support Vector Machine
(SVM) based approach to predict the location as alternative
to existing modeling approaches. SVM forms the new generation of machine learning techniques used to find optimal
separability between classes within datasets. Experiments
on two different spatial datasets show that SVMs gives reasonable results.

1 Introduction
Data mining employs algorithms and techniques from
statistics, machine learning, artificial intelligence, databases
and data warehousing, etc [1]. Data mining techniques have
been successfully applied to analyze spatial data, i.e., data
related to objects that occupy space. Spatial data carries
topological and/or distance information; it is often organized by spatial index structures and accessed by spatial access methods. These distinct features of a spatial database
bring challenges and opportunities for mining knowledge
from spatial data [2]. Spatial Data mining is a subfield of
data mining; it is a process that uses a variety of data analysis tools to discover spatial patterns and relationships in spatial data that may be used to make valid predictions [3, 4].
This has wide applications in Geographic Information Systems (GIS), remote sensing, image databases exploration,
medical imaging, robot navigation, and other areas where
spatial data are used. The main methods for spatial data
analysis include: spatial association rules extraction, clustering, and classification.
Data objects stored in a database are identified by their
attributes. Classification finds a set of rules which determine the class of each classified object according to its attributes. Objects with similar attribute values are classified

1-4244-0212-3/06/$20.00/2006 IEEE

952

ordinates or Latitude and Longitude. Also each location

is identified by its crime rate and several related attributes.
We develop a model to automatically classify the locations
as either hotspot crime members or non-members. Our system accepts a predefined level of crime rate as input. Based
on this value, the system will label a certain portion of the
dataset as hotspot members and non-members. If a locations crime rate is above the predefined level of crime, it
will be labeled as a member of hotspot crime class or positive sample, otherwise a non-member of hotspot crime or
negative sample. The user specifies from the dataset a certain portion to be used as input to the system.
We are using two different approaches for choosing the
certain portion of the dataset. After labeling the specified
certain percentage of the dataset, the system uses the labeled portion as a training set for building a classifier by
applying SVMs. Since we are classifying the locations into
positive and negative samples, we can approach this task as
a binary classification problem. The system automatically
classifies a location with crime rate above (below) the predefined level as positive (negative).
SVMs receive great consideration because of their attractive performance in a wide variety of application domains such as object recognition, speaker identification,
face detection, handwriting recognition and text categorization [12]. Generally, SVM solves classification problems
by learning from examples. As it is obvious from its name,
the binary classification method requires negative and positive examples to establish a statistical relationship and to
build a classifier model. However, in reality, many types
of datasets suffer from lack of reliable negative examples.
This is our main motivation to extend this problem to crime
classification. SVM have been recently applied in several
studies such as gene expression classification [13], text categorization [14] and text summarization [15].
To demonstrate the effectiveness and applicability of the
proposed approach, we report in the paper test results on
two datasets downloaded from the internet.
The rest of the paper is organized as follows. Section 2
covers the SVM information required for this study. Experiment methods are discussed in Section 3. The datasets utilized in the experiments are described in Section 4. Detailed
explanation on the selected evaluation model, experiments
and the results are provided in Section 5 and Section 6, respectively. Section 7 is summary and conclusions.

Figure 1. Definition of hyper-plane and margin: circular and square dots are samples of
classes -1 and +1, respectively

and distinguish whether each data instance belongs to positive or negative according to the Optimal Separating Hyperplane.
For most real-world problems that seem not to be linearly separable, SVMs can work in combination with the
kernel function technique [11], which automatically realizes a non-linear mapping onto a feature space. The Optimal Separating Hyper-plane found by SVM is the feature
space that corresponds to a non-linear decision boundary in
the input space [11].

Figure 2. Hypersphere with the target data described by center a and radius R. Three objects on the boundary as the support vectors.
One object xi outside and has i > 0

2 Support Vector Machine

2.1 One-Class SVM
SVMs perform classification, i.e., separate a given
known set of +1,-1 labeled training data via a hyper-plane
that is maximally distant from the positive samples and negative samples (Optimal Separating Hyper-plane as in Figure 1); plot the test data at the high dimensional space;

There are two different approaches to one-class SVMs.

The goal of the one-class SVM approach of Tax and
Duin [17], which is called the Support Vector Data Description (SVDD), is to find a hypersphere that covers as many

953

training data points as possible, while keeping the radius of

the hypersphere as small as possible. In other words, given
l training data points, xi , (i = 1, 2, ..., l), find the smallest
possible hypersphere to contain the training points in multidimensional space. As shown in Figure 2, small portion of
outliers are allowed to exist using a slack variable (i ) [18]:
1 X
i
vl i

Min(R2 ) +

(1)

Subject to:
(xi c)T (xi c) R2 + i , i 0, i [1, l]
where c and R are the center and radius of the sphere, respectively, T is the transpose, and v (0, 1] is the tradeoff
between volume of the sphere and the number of training
data points rejected. When v is large, the volume of the
sphere is small; so more training points will be rejected.
This optimization problem can be solved by the Lagrangian:
L(R, , c, ai , i ) =

1
vl

R2 +

l
P
i=1

l
P
i=1

Figure 3. The solid line is the conventional

classifier to distinguish between positives
and negatives, while the dashed line describes the dataset.

ai {R2 + i

(x2i 2cxi + c2 )}

l
P
i=1

i i

(2)

where ai 0 and i 0. Setting to zero the partial derivative of L with respect to R, ai , and c, we get:
l
X

ai = 1

(3)

1
vl

(4)

ai xi

(5)

i=1

0 ai
c=

l
X

where S and S are simple subsets of the input space and its
complement, respectively. Let : <N F be a nonlinear
mapping that maps the training data from <N to a feature
space F . To separate the data set from the origin, solve the
following primal optimization problem [15]:

i=1

Substituting Eqs. (3)-(5) in Eq. (2) gives the dual problem:

X
X
ai (xi .xj )
(6)
min
ai aj (xi .xj )
a

i,j

Subject to:
0 ai

1
vl ,

l
P
i=1

be used in order to build more flexible model that can deal

with non-linear data. Kernel functions and their usage in
building powerful classifiers are described in Section 2.2.
The second approach in one-class SVMs was proposed
by Scholkopf et al [16]. Their approach is to construct a hyperplane that is maximally distant from the origin with all
data lying on one side from the origin as shown in Figure 3.
In other words, given training set x1 , . . . ,xl RN , where
xi is a feature vector, it is required to estimate a function
that takes the value +1 in a small region capturing most of
the data points, and -1 elsewhere [14]. Formally, the function is written as:

+1 if x S
f (x) =
(8)
1 if x S

l
1
1 X
2
Minimize V (w, , ) = kwk +
i
2
vl i=1

ai = 1

Calculating the distance between a test point (x) and the

center C of the hypersphere determines whether (x) is inside the sphere or not. By using the following inequality,
the position of the test point can be identified:
X
X
(x.x) 2
ai (x.xi ) +
ai aj (xi .xj ) R2
(7)

(9)

Subject to:
(w.(xi )) i , i 0,
where v (0, 1) is a parameter for controlling the tradeoff between the number of outliers and model complexity,
and is the margin. Using the following decision function,
a label can be assigned to a new given data point (x) for
classification task:

In reality the data points are not always spherically distributed. So, different types of kernel functions K(xi , xj ) can

f (x) = sgn(w.(xi ) )

954

(10)

Again same as the former approach, introducing Lagrange

multipliers i and using the Kuhn-Tucker condition, the
derivatives with respect to the primal variables are set to
zero to get:
X
w=
i (xi )
(11)

satisfy Mercers conditions. Since two types of these functions are often used for classification problems; polynomial
and Gaussian kernels; here we limit ourselves to these two
kernel functions.
A polynomial mapping is a popular method for nonlinear modeling:

where only a subset of points xi that are closest to the hyperplane have nonzero values i . These points are called
support vectors. Instead of solving the primal optimization
problem directly, the following dual problem can be considered:
1X
Maximize w() =
i j K(xi , xj )
(12)
i,j
2
Subject to:
0 i

1
vl ,

P
i

K(x, x0 ) = hx, x0 i

K(x, x0 ) = (hx, x0 i + 1)

i = 1

K (xi , xj ) = ((xi ), (xj )) are the kernels (the dot products between mapped pairs of input points). Kernel functions allow more general decision functions when the data
points are not linearly separable.

3 Experiment Methods
In order to select a certain representative portion of the
crime datasets to be used as the training set by the system,
we experiment the following approach: for a given percentage of the data and a predefined level of crime rate, we select a subset of the crime dataset to label; and then based on
the predefined level of crime rate, we specify a class label
to each data point in the selected set. The data points which
have the crime rate above the predefined rate are positive
or members of hotspot class and data points with crime rate
below the predefined rate are negative or non-members of
hotspot class. Then this labeled data set will be used as the
training set in SVM classification. To select a given percentage of the data to be labeled, we use the k-median clustering
algorithm. Then, we compare the result when the same percentage of the data is selected randomly.

2.2 Kernel Functions

The idea of the kernel function is to enable operations
to be performed in the feature space rather than the potentially high dimensional input space. Hence, the inner product does not need to be evaluated in the feature space. This
provides a way of addressing the curse of dimensionality.
However, the computation is still critically dependent upon
the number of training patterns, and to provide a good data
distribution for a high dimensional problem will generally
require a large training set [19].
The kernel theory is based upon Reproducing Kernel
Hilbert Spaces (RKHS) [20, 21]. An inner product in
the feature space has an equivalent kernel in the input
space, K(x, x0 ) = ((x), (x0 )), provided certain conditions hold. If K is a symmetric positive definite function,
which satisfies Mercers Conditions,
am m (x)m , am 0

(14)

K(x, x0 )g(x)g(x0 )dxdx0 > 0, g L2

(15)

(17)

Classical techniques utilizing radial basis functions employ

some method of determining a subset of centers. Typically,
a method of clustering is first employed to select a subset
of centers. An attractive feature of the SVM is that this
selection is implicit, with each support vector contributing
one local Gaussian function, centered at that data point.

The second kernel is usually preferable as it avoids problems with the Hessian becoming zero.
Radial basis functions have received significant attention, most commonly with a Gaussian of the form:

!
0 2
kx

x
k
K (x, x0 ) = exp
(18)
2 2

From Eqs. (10) and (11), the decision function can be

eventually written as:

!
X
i K(xi , xj )
(13)
f (x) = sgn

K(x, x0 ) =

(16)

3.1 K-Mean Clustering Algorithm

K-means clustering is a partitioning method that partitions the data points of the input dataset into k clusters.
Each data point in the dataset is treated by k-means algorithm as an object having a location in space. The algorithms finds a partition in which objects within each cluster
are as close to each other as possible, and as far from objects in other clusters as possible. Each cluster is defined
by its data members and its centroid. The centroid for each

Z Z

then the kernel represents a legitimate inner product in the

feature space. There are many kinds of valid functions that

955

cluster is the point to which the sum of distances from all

objects in that cluster is minimized.
In our experiment, the idea of using k-means clustering
is to consider the data points within a certain distance from
the centroids of the k clusters as representative data points
of the clusters. The size of k should be chosen in a way by
the system such that a certain total percentage of the data
set falls into a desired distance from a closest cluster center.
By gathering a small set of data points from each cluster
based on the above procedure, we will have a portion of the
whole data set ready to be labeled as the training set. Points
having their crime rate above the predefined level of crime
are labeled as members of hotspot class and points having
crime rate less than the predefined level are labeled as nonmembers of hotspot class.
To compare the clustering approach and random approach for choosing a portion of the data set as the training
set, we devise the following experimental plans:

of crime hotspot class. Hotspot crime locations are places

which have the crime rate above the predefined crime level.
The second dataset [23] records 78 counties surrounding St. Louis, MO homicides rate and related variables
(see Figure 4c and 4d). The problem here is to distinguish
between members and non-members of homicide hotspot
class. Hotspot homicide locations are places which have
the homicide rate above the predefined level of homicide.

5 Model Evaluation
Basically, n-fold cross validation is a method in which
the data is randomly divided into n disjoint groups [24]. For
example, suppose the data is divided into ten groups. The
first group is set aside for testing and the other nine are put
together for model building. The model built on the 90%
group is then used to predict the group that was set aside.
This process is repeated a total of 10 times as each group
in turn is set aside. Finally, a model is built using all the
data. The mean of the 10 independent error rate predictions
is used as the error rate for this final model. In our study, a
five-fold cross validation method has been used to estimate
the accuracy of the classification model.

1. Random Selection for Labeling + One-Class SVM:

we randomly select a given percentage of the data
points as a small representative portion of the crime
dataset to be labeled. Then, the one-class SVM algorithm uses the output labeled set as the training set to
build the classifier.

6 Experiment and Result

2. Clustering-based Selection for Labeling + OneClass SVM: k-means is used to select a given percentage of the data points to be labeled. Then, the oneclass SVM algorithm uses the output labeled set as the
training set to build the classifier.

For our experiments we used Personal Computer with

Intel P4 2.4GHZ CPU and 1GB memory. The experiments were carried out by using a Matlab interface of LIBSVM [25] in Matlab 7. LIBSVM is a library for SVM classification and regression. The predefined level of crime depends on the knowledge of domain expert and is usually
specified by crime experts. In this experiment, we assume
that the data points of our datasets follow a normal distribution (Gaussian distribution) so that the optimal average
value of the crime C is halfway between Cmin and Cmax .
According to the definition of the Gaussian distribution Caverage and Cmax are the mean (average) and the
mean incremented by standard deviation (variability), respectively. In our first set of experiments, we considered
Caverage as the predefined level of crime rate, i.e., if the
crime rate of a sample location is above Caverage , then that
location is a member of hotspot class; otherwise it would be
a non-member of the hotspot class. In the second set of the
experiments, Cmax was considered as the predefined value
for hotspots crime rate.
We selected 20% of the datasets to be labeled and used
as training data in the SVM algorithm. In our experiments,
we also take into account the case in which we assign the
remaining 80% unlabeled data to negative class or hotspot
non-member class, and give the complete labeled data set to
the SVM to be trained. Then, we compare the result with the

3. Complete Data Set + One-Class SVM: after choosing a certain percentage of the dataset for labeling by
random selection or clustering technique, we label the
rest of the dataset as negative samples and add them to
the training set. Then, we pass the complete labeled
dataset to one-class SVM as the training set.

4 Datasets
To test the different approaches used by our model, we
downloaded two published crime datasets from the internet. The datasets consist of the crime rate and related variables for each data point. The location of each data point
is described in the dataset by Euclidian coordinates or Latitude and Longitude. The datasets were downloaded from
[http://www.terraseer.com/]. In this section, we will provide
short description for each dataset. For further information,
please, refer to [http://www.terraseer.com/].
The first dataset is a small crime dataset [22] that records
crime rate and 20 related variables in 49 neighborhoods in
Columbus Ohio, USA (See Figure 4a and 4b). The problem is to distinguish between members and non-members

956

Figure 4. (a) Crime data and (b) its rate distribution; (c) Homicide data and (d) its rate distribution.

case where we ignore the remaining 80% unlabeled data. In

our first approach to data selection, which is random data
selection, we selected 20% of the dataset randomly. In order to get a more consistent result, we performed this experiment 20 times.

Table 1. 20% of the dataset selected randomly

and labeled.
Data Set

In the clustering-based data selection approach, we applied k-mean clustering algorithm to the datasets first. This
second approach of data selection chooses the data more
wisely than the random selection. After labeling the data,
we used the result set as input to the SVM algorithm. In
one-class SVMs, we applied different kernel functions to
see how it would influence the classification accuracy. The
results for four different experiments are shown in Tables 14. As it can be seen in each experiment, first we evaluate the
effect of using different kernel function for one-class SVM
technique. Here we have chosen Linear, Polynomial and
Gaussian kernel functions. We applied the default values of
LIBSVM for Polynomial and Gaussian. We also changed
the parameter in a range to see whether the result will be
improved or not. We also performed each step 20 times and
presented the average of 20 runs as the final result.

Columbus
St. Louis

C
35.13
51.86
4.57
10.58

Linear
63.00
58.50
53.13
50.63

One-Class SVM
Polynomial Gaussian
60.50
68.50
57.50
62.50
53.44
49.69
50.94
48.44

dataset is selected randomly and labeled. This portion of

data is passed to SVM as the training set. Then, we run the
SVM classifier with different kernel functions on the data
set; Gaussian kernel performs better. Based on our knowledge from the Columbus dataset, when we set the predefined crime rate to mean value, number of positive samples
will be more than the case when we set the predefined value
C to mean plus standard deviation. Therefore, we can say
that one-class SVM performs better when we have a small
training set with more positive samples.
In Table 2, the result is displayed when 20% of the
dataset is selected randomly and labeled, and the rest is labeled as negative samples and added to the training set. The
difference between this experiment and the one shown in
Table 1 is that we have increased the size of the training set
by adding more negative samples. As it can be seen in Table 2, one-class SVM will result in a noticeably decreasing

All the results from different techniques are expressed as

the percentages of correctly classified samples on the test
data. The values of C are set to the mean () and the mean
plus the standard deviation ( + s) calculated for each experimented dataset.
In Table 1, the result is displayed when 20% of the

957

Table 2. 20% of the dataset selected randomly

and labeled; the rest 80% labeled as negative
samples and added to the training set.
Data Set
Columbus
St. Louis

C
35.13
51.86
4.57
10.58

Linear
52.45
51.22
49.94
50.39

Table 4. 20% of the dataset selected by clustering and labeled; the rest 80% labeled as
negative samples and added to the training
set.

One-Class SVM
Polynomial Gaussian
51.43
57.04
51.53
55.10
49.62
38.65
50.06
38.46

Data Set
Columbus
St. Louis

classification performance as we have a larger training set

with more negative samples.

Columbus
St. Louis

C
35.13
51.86
4.57
10.58

Linear
63.89
78.33
53.44
51.25

35.13
51.86
4.57
10.58

Linear
53.47
53.37
51.47
50.51

One-Class SVM
Polynomial Gaussian
53.06
57.35
53.37
57.76
51.09
39.23
50.64
39.31

advantages. First, one-class SVMs construct a hyperplane

that is maximally distant from the origin with all data lying on the opposite side from the origin. This helps to find
an optimal hypersphere which contains all or most of the
training points which belong to the positive samples. Furthermore, using different types of kernel functions gives an
ability to one-class SVMs so that it represents various data
distribution shapes in feature space (e.g., spherical shapes
or very irregular shapes). Second, one-class SVMs make no
assumption on the probability density of the data. This is a
useful benefit when the data do not follow any probability
distribution (such as a normal distribution), or insufficient
data are available to test the distribution [18].

Table 3. 20% of the dataset selected by clustering and labeled.

Data Set

One-Class SVM
Polynomial Gaussian
61.11
69.44
76.11
65.55
54.38
50.94
51.56
24.38

In Table 3, the result is displayed when 20% of the

dataset is selected by clustering and labeled. Then this portion of data is passed to SVM as the training set and we
run the SVM classifier with different kernel functions on
the data set. Compared to Table 1, one-class SVM performs
much better when we use clustering instead of random selection. As a matter of fact, by clustering technique we
choose the data more wisely than we do in random selection. The size of k will be chosen by the system such that
we will be able to select 20% of the original data set from
the different clusters. This means we will select 20% of
the dataset within a certain distance from the closest center
of each cluster and then we put them all together and label
them as the training set. Same as our first experiment, we
see that one-class SVM with Linear and Gaussian kernels
performs better when C is set to the mean value of crime
rate, which means we have a small training set with more
positive samples.
Based on the results from our previous experiments, we
can predict the expected result from our last experiment
where 20% of the dataset is selected by clustering and labeled and the rest 80% is labeled as negative sample and
added to the training set. The result is displayed in Table 4.
One-class SVM performs better compared to Table 2, since
we are using clustering instead of random selection. However, it will do a poor job compare to Table 3, where we
have a smaller dataset with more positive sample than we
do in Table 4.
To sum up, one-class SVMs is attractive for two main

7 Summary and Conclusions

In this study, we provided an inquiry system which can
be used as a general framework in different domains and by
domain experts in order to customize the spatial data classification task. In our framework, we focused on some special types of spatial datasets like the ones we used for our
experiments. As a case study in the area of public safety,
we concentrated on the performance of one-class SVMs
for predicting the hotspot crime location when a predefined
level of crime rate and a percentage for selecting a portion of that are given. We applied two different approaches
for data selection, first we chose the certain portion of the
data randomly and in the second approach we applied kmeans clustering algorithm in order to make a more wise
selection. Then we labeled the selected portion of the data
set as members and non-members of crime hotspot class
based on the predefined level of crime rate. We also studied the case when the rest of the dataset were labeled as
non-members samples and added to the training set. Our
experiments show that one-class SVM gives reasonable result when we choose appropriate parameters for the algorithm. Based on our different experiences shown in the result tables, we can conclude that SVMs form an appropriate
approach to hotspot crime prediction, while k-means clustering algorithm is useful for data selection and by using
the rest of the dataset as non-member samples. This moti-

958

vates us to investigate the same problem by using two-class

SVMs; we expect to achieve much better results. Our plan
is to compare one-class and two-class SVMs with the other
approaches to hotspot crime prediction.

[12] N. Cristianini and J. S. Taylor, An Introduction to Support Vector Machines, Cambridge University Press,
2000.
[13] A. Kowalczyk and B. Raskutti, One Class SVM for
Yeast Regulation Prediction, SIGKDD Explorations,
Vol.4, pp.99-100, 2002.

References
[1] M.H. Dunham, Data Mining: Introductory and Advanced Topics, Prentice Hall, 2002.

[14] L. M. Manevitz and M. Yousef, One-Class SVMs for

Document Classifcation, Journal of Machine Learning, 2, pp.139-154, 2001.

[2] W. Lu, J. Han, and B. C. Ooi, Discovery of General

Knowledge in Large Spatial Databases, Proc. of Far
East Workshop on Geographic Information Systems,
pp.275-289, Singapore, June 1993.

[15] C. Kruengkrai and C. Jaruskulchai, Using One-Class

SVMs for Relevant Sentence Extraction, Proc. of the
International Symposium on Communications and Information Technologies, 2003.

[3] K. Koperski, J. Adhikary, and J. Han, Spatial Data

Mining: Progress and challenges, Proc. of ACM SIGMOD Workshop on Research Issues on Data Mining
and Knowledge Discovery, pp.55-70, Montreal, 1996.

[16] B. Scholkopf, et al, Estimating The Support of

a High Dimensional Distribution, Technical report,
Microsoft Research, MSRTR9987, 1999.
[17] D. M. Tax and R. P. Duin, Outliers and Data Descriptions, Proc. of the Annual Conference of the Advanced School for Computing and Imaging, 2001.

[4] K. Koperski and J. Han, Discovery of Spatial Association Rules in Geographic Information Databases,
Proc. of the international Symp. on Large Spatial
Databases, pp.47-66, Portland, Maine, 1995.

[18] Q. Gou, M. Kelley and C. H. Graham, One-Class

Support Vector Machines for Predicting Distribution
of Sudden Oak Death in California, Ecological Modeling, 182:75-90, 2005.

[5] S. Shekhar, et al, Trend in Spatial Data Mining, to

appear in Data Mining: Next Generation Challenges
and Future Directions, in H. Kargupta, A. Joshi, K.
Sivakumar, Y. Yesha (eds.), AAAI/MIT Press, 2003.

[19] V.N. Vapnik, The Nature of Statistical Learning Theory, Second Edition, Springer, New York, 1999.

[6] S. Ozesmi and U. Ozesmi, An Artificial Neural

Network Approach to Spatial Habitat Modeling
with Interspecific Interaction, Ecological Modeling,
(116):15-31, 1999.

[20] N. Aronszajn, Theory of Reproducing Kernels,

Trans. Am. Math. Soc., Vol.68, pp.337-404, 1950.
[21] G. Wahba, Spline Models for Observational Data,
Proc. of SIAM CBMS-NSF Regional Conference Series in Applied Mathematics, v. 59, 1990.

[7] U. Ozesmi and W. Mitsch, A Spatial Habitat Model

for Marsh-Breeding Red-Winged Black Bird, Ecological Modeling, (101):139-152, 1997.

[22] L. Anselin, Spatial Econometrics: Methods and Models, Dordrecht: Kluwer Academic, Table 12.1 p. 189,
1998.

[8] S. Chawla, S. Shekhar, W. Wu, Predicting Locations Using Map Similarity (PLUMS): A Framework
for Spatial Data Mining, Proc. of ACM International
Conference on Knowledge Discovery and Data Mining, Boston, MA, 2000.

[23] S. Messner, et al, The Spatial Patterning of County

Homicide Rates: An Application of Exploratory Spatial Data Analysis, Journal of Quantitative Criminology, Vol.15, No.4, pp.423-450, 1999.

[9] J. Lesage, Regression Analysis of Spatial Data, The

Journal of Regional Analysis and Policy, Vol.27, No.2,
pp.83-94, 1997.

[24] H. A. Edelstein, Introduction to Data Mining and

Knowledge Discovery, Third Edition, Potomac, MD:
Two Crows Corp, 1999.

[10] B. E. Boser, I. M. Guyon and V. Vapnik, A Training Algorithm for Optimum Margin Classifiers, Proc.
of the Annual Workshop on Computational Learning
Theory, Pittsburgh. ACM, 1992.

[25] C. C. Chang and C. J. Lin, LIBSVM: A Library for Support Vector Machines, 2001. URL:
http://www.csie.ntu.edu.tw/ cjlin/libsvm.
[26] J. P. LeSage, MATLAB Toolbox for Spatial
Econometrics, 1999. URL: http://www.spatialeconometrics.com.

[11] V. Vapnik, Statistical Learning Theory, John Wiley,

NY, p.732, 1998.

959