Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
KEY WORDS: hot-spot detection; crime; cluster analysis; spatial analysis; geo-
graphic information system (GIS).
1. INTRODUCTION
The ‘‘art’’ and science of cluster analysis continues to reward and
frustrate criminologists, geographers, sociologists, and others in their efforts
to apply this statistical technique in crime mapping and analysis. Broadly
1
Department of Geography, University of Cincinnati, Cincinnati, OH, 45221-0131, USA;
Phone: +1-513-556-3357; E-mail: tony.grubesic@uc.edu
77
0748-4518/06/0300-0077/0 2006 Springer Science+Business Media, Inc.
78 Tony H. Grubesic
2
Additional discussions of crime hot-spots can be found in Ratcliffe and McCullagh (2001),
Craglia et al. (2000, 2001), and Ackerman and Murray (2004).
Application of Fuzzy Clustering for Crime Hot Spot Detection 79
3
Nearest neighbor measures are typically based on Euclidean distances.
Application of Fuzzy Clustering for Crime Hot Spot Detection 81
i = index of observations;
ai = attribute weight of observation i (i.e. the number of crimes at a given
location);
k = index of clusters;
dik= distance between observation i and cluster k;
n
1 if observation i is assigned in cluster k
yik ¼
0 otherwise
K-means
XX
Minimize Z ¼ ai d2ik yik ð1Þ
i k
Subject to
X
yik ¼ 1 ð2Þ
k
perspective, such biases can mislead and distort the assignment of resources
(e.g. police patrols). Given the budget constraints that most police depart-
ments are facing, this is a significant concern. Further, efforts in crime
forecasting (Gorr and Harries, 2003) or prospective hot-spotting (Bowers
et al., 2004) must be built on a foundation of strong, unbiased empirical
results, stemming from recursive crime analysis. This clearly reiterates the
need for a careful evaluation and choice of statistical techniques, particu-
larly when using spatial data for analysis. Moreover, it underscores the
importance of accurate hot-spot generation and interpretation.
Given the need for spatial accuracy, and the problems associated with
the distance metric in the k-means heuristic, there is a need to implement a
clustering algorithm that is not geographically biased to support hot-spot
detection and delineation. A viable option is the median clustering algo-
rithm. This approach differs from the both the k-means and central point
(Murray, 1999) clustering approach in several important ways. First, instead
of using an artificial point, such as the centroid generated in the k-means
procedure, to identify spatial clusters, the median approach uses the actual
observations (e.g. crime events), to identify clusters (Hakimi, 1964; Murray
and Estivill-Castro, 1998; Vinod, 1969). The following notation will be
utilized in the specification of this alternative approach:
i = index of observations;
j = index of potential medians;
ai = attribute weight of observation i (i.e. the number of crimes at a given
location);
dij = distance between observation i and potential median j;
p = number of cluster medians to be selected;
n
1 if cluster median j is selected
xj ¼
0 otherwise
n
1 if observation i is assigned to cluster median j
zij ¼
0 otherwise
Median clustering problem (MCP)
XX
Minimize Z ¼ ai dij zij ð4Þ
i j
Subject to
X
zij ¼ 1 for all i ð5Þ
j
Application of Fuzzy Clustering for Crime Hot Spot Detection 85
X
xj ¼ p ð7Þ
j
xj ¼ ð0; 1Þ
From a functional standpoint, this approach is virtually identical to the
partitioning around medoids (PAM) algorithm outlined by Kaufman and
Rousseeuw (1990). The objective (4) of the MCP is to minimize the total
weighted assignment of observations to selected medians. This, in effect,
minimizes the average distance between observations and their assigned
median. Constraints (5) force all observations to be assigned to a median.
Constraints (6) require that a median be selected before it serves as a rep-
resentative location for grouping observations. Constraint (7) specifies that
p clusters are identified and constraint (8) imposes integer restrictions on the
decision variables.
It is important to note that the objective function of this model is linear,
which provides significant computation advantages over the k-means
approach (Murray and Estivill-Castro, 1998). Second, the use of a Euclid-
ean distance metric in the median model provides two superior features. Not
only can the distances be calculated prior to running the model, the
Euclidean metric also provides a real physical interpretation of distance,
representing travel, transportation or movement. This more realistic rep-
resentation of geographic space is critical for spatial cluster detection and
hot-spot delineation.
3. FUZZY CLUSTERING
As mentioned previously, fuzzy clustering is a generalization of parti-
tioning. Instead of observations being assigned to one and only one cluster,
the fuzzy approach allows for some ambiguity in the data. That is, obser-
vations have a degree of belonging to one or more clusters (Hoppner et al.,
1999). In this way, cluster membership for each observation is spread out
over a range of groups. Fuzzy clustering is used in a wide variety of fields,
including finance (Boreiko, 2003), animal science (Chen et al., 2004), agri-
cultural risk assessment (Zhang, 2004) and meteorology (Liu and George,
2003). Kaufman and Rousseeuw (1990) note that the main advantage of
fuzzy clustering over strictly partition-based methods is that the fuzzy ap-
proach yields much more detailed information on the structure of the data.
As illustrated in Fig. 2, if a data set contains intermediate cases (i.e. inter-
mediate spatial locations) between relatively solid groups, the fuzzy ap-
proach is much better equipped to deal with this structure. The following
notation will be used to specify the fuzzy approach:
Subject to
5
These properties will be explored in the next section.
88 Tony H. Grubesic
1 2 6 7
5
3 4 8 9
Cluster 1 Cluster 2
10 11
X
uik ¼ 1 for all i ð10Þ
k
Not surprisingly, some fuzzy clusters are fuzzier than others. If observations
have equal memberships in all clusters, (1/k), the groupings are completely
fuzzy. If the observations have a value of 1 for only one cluster each, the
groupings are completely hard (partition based). Therefore, in order to
measure the degree of fuzziness in cluster groupings, one can implement
Dunn’s partition coefficient (1976), which is expressed by:
XX
Fk ¼ u2ik =n ð12Þ
i k
Expressed differently, this simply means that for a completely fuzzy clus-
tering, all uik will take a minimal value of 1/k. A completely hard partition
yields the maximal value of Fk=1. There are several important points worth
noting in this model. First, the FCP is a generalized version of the MCP
(Kaufman and Rousseeuw, 1990). Therefore, this approach is still
attempting to minimize the aggregate distance between observations and
potential medians when generating clusters.7 Second, the FCP is not for-
mulated explicitly as a location model, as was done for the k-means and
MCP in this paper. The basic difference is that a parameter for weighting
observations is not included in the objective function. The FCP is simply
concerned with minimizing the measure of dissimilarity, which in this case,
corresponds to the distance metric.
The following section explores the differences in these three approaches
to cluster detection, focusing on the utility of a fuzzy approach for hot-spot
detection. This includes both a functional and graphical comparison of these
clustering methods. A method for determining the relative strength of
cluster composition will also be highlighted.
4. EMPIRICAL RESULTS
Comparative results for the three clustering approaches will be pre-
sented for a spatial application across a range of k cluster values using a
Pentium 4/2.53 GHz computer. Application data contain 613 crime events
from the Pleasant Ridge neighborhood of Cincinnati, Ohio, for the year
2003. These data were acquired from the Cincinnati Police Department and
geocoded using Centrus Desktop 4.0. ArcGIS version 8.2 was implemented
to manage, manipulate and analyze the crime observations. NCSS 2004 was
utilized for all statistical calculations, including the clustering approaches.8
7
The FCP can be solved using iterative approach that stops when the objective function con-
verges (Kaufman and Rousseeuw, 1990).
8
NCSS (http://www.ncss.com) limits the sample size for both the MCP and FCP to 1000
observations.
90 Tony H. Grubesic
Figure 3 illustrates the study area for this analysis. All geographic data
layers are projected to the Ohio State Plane coordinate system using North
American Datum 1983. The units are measured in feet. Boundary polygons
correspond to a specific police district, beat and zone for Cincinnati. In this
case, the Pleasant Ridge neighborhood is located approximately 8 miles
northeast of the central business district in District 2, Beat 3, Zone 1. As
mentioned previously, there were 613 crime events in this area during 2003,
ranging from petit theft to murder.
4.1. k-Means
Cluster model solutions for the k-means approach are displayed in
Fig. 4. The column representing ‘‘percent variation’’ in Fig. 4 provides the
within-cluster sum of squares for k, as a percentage of the within sum of
squares with no clustering. In other words, lower values suggest less spatial
variation between members in the cluster groups. This is reflected in the
gradual decline of the variance in Fig. 4; as the number of k (groups)
60 60
Change
Variation
50 50
40 40
Percent
30 30
20 20
10 10
0 0
2 3 4 5 6 7 8 9 10 11 12 13 14
k
Change Variation
Hull 3 (433.33)
Hull 4 (775.36)
Hull 1 (1222.22)
Hull 5 (1762.37)
Hull 6 (371.42)
Hull 2 (612.36)
Hull 8 (580.35)
( ) Hull Density
Hull 7 (285.71)
are both informative and easy to calculate. One simply divides the area of a
convex hull by the total number of points enclosed by that same hull. Higher
density ratios suggest elevated crime rates while lower density ratios suggest
the opposite.
In the case of Fig. 5, simple geometric calculations reveal that both
Hull 1 (i.e. Cluster 1) and Hull 5 (i.e. Cluster 5) are areas of elevated crime.
Hull densities are 1222 and 1762 crimes per square mile, respectively. It is
interesting to note that the spatial distribution of these crimes and the areas
defined by each of these hulls are quite different. For example, Hull 1 is
more linear in nature, following a major commercial thoroughfare
(Losantiville Road) while Hull 5 constitutes a much greater expanse,
incorporating several major roads and commercial centers. Cartographic
investigation suggests several other areas where concentrations of crime
exist. For example, both Hull 2 and Hull 4 have relatively dense cores of
crime events. However, the cluster groups and convex hulls encompass a
relatively large area, making hull densities (612 and 775 crimes per square
mile, respectively) much lower.
This, in summary, is what makes the all-or-nothing assignment of crime
events in non-hierarchical cluster analysis problematic for hot-spot detec-
tion and analysis. In the case of Hull 4, because each crime event must be
Application of Fuzzy Clustering for Crime Hot Spot Detection 93
assigned to one and only one cluster, outliers and intermediate cases can
both skew and hide the true spatial composition of hot-spots.
4.2. MCP
Figure 6 displays the results the median (medoid) based approach
outlined in Section 2, which utilizes a standard Euclidean distance measure.
Specifically, Fig. 6 displays the diagnostic statistics for this approach, which
includes the weighted distance and silhouette values for a series of k values.
Silhouettes are both graphical and quantitative diagnostics for cluster
memberships introduced by Rousseeuw (1987). Each cluster in the median/
medoid based approach is represented by a single silhouette. Silhouettes are
created by measuring the average dissimilarity of each object in a cluster
group to all other objects in the cluster group. The silhouette values s(i)
representing the membership of an event to its assigned cluster are calcu-
lated for each object i. The value for s(i) always lies between )1 and 1. This
value may be interpreted as follows:
sðiÞ 1 ) object i is well classified.
sðiÞ 0 ) object i is an intermediate case.
sðiÞ 1 ) object i is poorly classified
0.5
0.45
Average Dissimilarity
0.4
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0
4 5 6 7 8 9 10 11 12 13 14 15
k
Normalized Dunn’s
Number of clusters Average distance Average Silhouette partition coefficient (Fk)
Both the individual values and average values can be plotted for visual
comparison (see Kaufman and Rousseeuw, 1990). The graphed values in
Fig. 6 suggest that k=8 (0.4599) is a good choice for revealing a relatively
natural grouping of the crime events. Results also suggest that k=14
(0.4656) might be another good choice. These grouping schemes represent
the two highest average silhouette values for the crime data (Fig. 6). For
illustrative purposes, k=8 will be utilized for a comparative analysis.
Figure 7 displays the spatial characteristics of this grouping. Although
the spatial differences between the k-means and MCP approaches are evi-
dent, there are still problems with the MCP solution for the delineation and
interpretation of crime hot-spots. The most significant hull densities still
belong to Hull 1 (1222) and Hull 5 (1453). However, similar to the k-means
approach, the median clustering model requires that each observation be
included in a group, with an all-or-nothing assignment. So, even though the
median approach accounts for geographic space in a much more realistic
fashion, spatial outliers must still be assigned to a cluster, regardless of
whether or not they actually contribute any interpretive value to a solution.
Perhaps the best examples are illustrated in Fig. 7, for Hull 7 and Hull 4. At
the far southern end of the Hull 7 grouping exists a series of five crime events
near an interstate highway exchange. These events drastically skew the
spatial composition of the convex hull. For example, if the five events in
question were completely removed from the analysis, the modified hull
would consist of 0.194 square miles (0.276 original) and a new hull density
of 376 versus 292 in the original. This represents a 22.34% change in density
value. Although these outliers create some interpretive difficulties in iden-
tifying hot-spots, one would not want to arbitrarily remove crime events in
Table II. FCP Output
Generalized
Event Offense Address FCP assignment Cluster 1 Cluster 2 Cluster 3 Cluster 4 Cluster 5 Cluster 6 Cluster 7 Cluster 8
1 Grand Theft 2920 Highland DR 8 0.1247 0.1076 0.1460 0.0419 0.0450 0.0773 0.0593 0.3981*
2 Breaking and 3467 Kimberly CT 2 0.0259 0.4047* 0.0408 0.0316 0.1623 0.1324 0.1556 0.0467
Entering
3 Breaking and 6011 Montgomery Rd 6 0.0080 0.0353 0.0155 0.0123 0.0119 0.8441* 0.0664 0.0064
Entering
4 Vehicle Theft 3154 Troy Av 7 0.0222 0.0601 0.0291 0.2139 0.2058 0.1553 0.2973* 0.0163
5 Criminal 6334 Montgomery Rd 5 0.0018 0.0088 0.0025 0.0058 0.9438* 0.0108 0.0246 0.0019
Damaging
Application of Fuzzy Clustering for Crime Hot Spot Detection
order to ‘‘cook’’ the data for a stronger result. This is why the all-or-nothing
assignment requirements of partition-based cluster analysis present a
problem. Every event must be evaluated and assigned to a cluster, regardless
of its geographic location or its potential contribution to hot-spot delinea-
tion. Partition-based algorithms are not equipped to handle spatial clusters
in a meaningful way, particularly where hot-spot detection and delineation
are concerned. Fortunately, the ability to handle outliers is a strength of the
FCP.
4.3. FCP
Table I displays the statistical diagnostics for the FCP solution. Similar
to the k-means and MCP approaches, the goal is to minimize within-group
variation and maximize between-group variation. Both the MCP and FCP
utilize average distance as the measure of within-group variation and the
average silhouette value as a measure of cluster quality. From an interpretive
Hull 3 (433.33)
Spatial Outlier
Hull 1 (1222.22)
Hull 4 (834.86)
Hull 5 (1453.70)
Hull 6 (389.38)
Hull 2 (639.05)
Hull 8 (540.14)
9
k=8 is clearly not the optimal solution for the fuzzy cluster analysis. However, this choice will
help readers make better and more informed visual comparisons between the FCP, k-means
and MCP results.
10
Similar to the MCP approach, silhouette values can be used to identify the strength/quality of
the derived clusters.
98 Tony H. Grubesic
Hull 3 (433.33)
Hull 4 (685.71)
Hull 8 (2250.00)
Hull 2 (756.75)
Hull 6 (438.91)
Hull 7 (278.91)
( ) Hull Density
increases. In this case, only values ranging between 32.1% and the maxi-
mum value of 95.5% are displayed. It is important to note that this surface
is not a kernel density map for a hot-spot analysis; it simply represents a
probability surface for membership in Cluster 5 based on the FCP.
Figure 10, however, utilizes the probability data for a more meaningful
interpretation of areas with elevated crime. Once again, the geometric
properties of the convex hull are leveraged for hot-spot delineation and
interpretation. Figure 10 displays two sets of convex hulls. The larger geo-
graphic set represents the generalized assignment of each observation with
the FCP algorithm. Specifically, these hulls bound each crime event to their
corresponding spatial clusters using the maximized probability value for
each observation. The crime density statistics for each of these hulls is
displayed in the ‘‘original density’’ column in Fig. 10. The second set of
convex hulls, which are clearly smaller in geographic extent, highlight the
differences in probability values as they range from 0 to 1 for each cluster. In
this case, only crime events with a probability value of 50% or higher for
their respective clusters are enclosed by the convex hull. This serves two
functions. First, it effectively eliminates spatial outliers in the analysis. In the
case of Hull 7, the five observations previously mentioned are no longer
Application of Fuzzy Clustering for Crime Hot Spot Detection 99
included in the modified (i.e. weighted) convex hull. The same can be said
for the spatial outlier that was highlighted in Hull 4 (Figure 7). This process
is one way to make an objective decision on the elimination of spatial
outliers for hot-spot detection. Instead of arbitrarily removing observations
to uncover areas of elevated crime density that were masked by the influence
of spatial outliers, one can actually make an objective decision based on the
calculated membership probabilities. Further inspection of Figure 10 re-
veals that many of the observations in the original hulls are no longer
included in the weighted versions. Clearly, this is a function of the cut-off
point of 50% for observation inclusion. If the probability cut-off point for
the weighted hulls was reduced to 30 or 40%, more observations would (and
could) be included. This level of flexibility is an attractive feature of the
FCP, providing the analyst an opportunity to make an informed decision on
which values make the best cut-off points. The second important function of
the weighted hulls is the ability to recalculate crime density and delineate
100 Tony H. Grubesic
Hull 3
Hull 4
Hull 1 Hull 5
Hull 8
Hull 2
Hull 6
Hull 7
Fig. 10. Compariation between original and weighted convex hulls for hot-spot analysis.
hot-spots. In this case, the new density values are in the second column of
the chart included in Figure 10, with the percent change highlighted in the
third column. The most significant changes occurred in Hulls 1 and 4. Hull 1
displays a 154% increase in crime density. Is this indicative of a hot-spot?
Not necessarily. The weighted density is still far lower than comparable
densities in Hulls 5 and 8. However, it is interesting to note the radical
change in Hull 4. A 101% increase, with a weighted density of 1379, clearly
indicates the presence of a crime hot-spot at the neighborhood level for
Pleasant Ridge. Not only is this the third highest value of all cluster groups
in the application, the dense core of crime events for Hull 4 was masked by
Application of Fuzzy Clustering for Crime Hot Spot Detection 101
REFERENCES
Ackerman, W. V., and Murray, A. T. (2004). Assessing spatial patterns of crime in Lima, Ohio.
Cities 21(5), 423–437.
Aldenderfer, M., and Blashfield, R. (1984). Cluster Analysis, Sage Publications, Beverly Hills.
Arabie, P., and Hubert, L. (1996). An Overview of Combinatorial Data Analysis: Clustering and
Classification, World Scientific Publishers, Singapore.
Bailey, T. C., and Gatrell, A. C. (1995). Interactive Spatial Data Analysis, Longman, Harlow.
Belbin, L. (1987). The use of non-hierarchical allocation methods for clustering large sets of
data. Aust. Comput. J. 19: 32–41.
Application of Fuzzy Clustering for Crime Hot Spot Detection 103
Bezdek, J. C. (1981). Pattern Recognition with Fuzzy Objective Function Algorithms, Plenum
Press, New York.
Boreiko, D. (2003). EMU and accession countries: Fuzzy cluster analysis of membership. Int. J.
Finance Econ. 8(4), 309–325.
Bowers, K. J., Johnson, S. D., and Pease, K. (2004). Prospective hot-spotting: The future of
crime mapping? Br. J. Criminol. 44: 641–658.
Brantingham, P. J., and Brantingham, P. L. (1981). Environmental Criminology, Sage Publi-
cations, Beverly Hills.
Brimicombe, A. J. (2003). A variable resolution approach to cluster discovery in spatial data
mining. In Kumar, V. et al. (eds.), ICCSA 2003.
Chainey S., and Cameron J. (2000). Understanding Hot Spots. Presentation prepared for the
2000 Crime Mapping Research Center Conference. San Diego, CA.
Chen, G. H., Wu, X. S., Wang, D. Q., Qin, J., Wu, S. L., Zhou, Q. L., Xie, E., Cheng, R., Xu,
Q., Liu, B., Zhang, X. Y., and Olowofeso, O. (2004). Cluster analysis of 12 Chinese native
chicken population using microsatellite markers. Asian-australas. J. Anim. Sci. 17(8), 1047–
1052.
Cohen, L. E., and Felson, M. (1979). Social change and crime rate trends: A routine activity
approach. Am. Sociol. Rev. 44: 588–607.
Craglia, M., Haining, R., and Wiles, P. (2000). A comparative evaluation of approaches to
urban crime pattern analysis. Urban Stud. 37(4), 711–729.
Craglia, M., Haining, R., and Signoretta, P. (2001). Modelling high-intensity crime areas in
English cities. Urban Stud. 38(11), 1921–1941.
Estivill-Castro V., and Murray A. T. (2000). Hybrid optimization for clustering in data mining.
CLAIO 2000 on CD-ROM, IMSIO, Mexico.
Fisher, W. (1958). On grouping for maximum homogeneity. J. Am. Stat. Assoc. 53: 789–798.
Gatrell, A. C., and Rowlingson, B. S. (1994). Spatial point process modeling in a geographical
information system environment. In Fotheringham, S., and Rogerson, P. (eds.), Spatial
Analysis and GIS, Taylor and Francis, London.
Gatrell, A. C., Baley, T. C., Diggle, P. J., and Rowlingson, B. S. (1996). Spatial point pattern
analysis and its application in geographical epidemiology. Trans. Inst. Br. Geogr. 21: 256–
274.
Gordon A. D. (1996). How Many Clusters? An investigation of five procedures for detecting
nested cluster structure. In Forer, P., Yeh A., and He, J. (eds.), Proceedings of 9th Inter-
national Symposium on Spatial Data Handling. Beijing International Geographical Union.
Gordon, A. D. (1999). Classification, Chapman and Hall, New York.
Gorr, W., and Harries, R. (2003). Introduction to crime forecasting. Int. J. Forecast. 19: 551–
555.
Graham, R. (1972). An efficient algorithm for determining the convex hull of a finite point set.
Info. Proc. Letters 1: 132–133.
Greenburg, S., and Rohe, W. (1984). Neighborhood design and crime. J. Am. Plann. Assoc. 50:
48–61.
Grubesic T. H., and Murray A. T. (2001). Detecting Hot Spots Using Cluster Analysis and GIS.
Fifth Annual International Crime Mapping Research Conference. Dallas, TX.
Grubesic, T. H., and Murray, A. T. (2002). Imperfect Spatial Information: Implications for
Crime Mapping and Analysis, Sixth Annual International Crime Mapping Research Con-
ference, Denver, CO.
Hakimi, S. L. (1964). Optimum locations of switching centers and the absolute centers and
medians of a graph. Oper. Res. 12: 450–459.
Hansen, P., and Jaumard, B. (1997). Cluster analysis and mathematical programming. Math.
Program. 79: 191–215.
104 Tony H. Grubesic
Harries, K. (1999). Mapping Crime: Principle and Practice, National Institute of Justice (NCJ
178919), Washington, DC.
Hintze J. (2001). NCSS and PASS. Number Cruncher Statistical Systems. Kaysville, Utah.
Hoppner, F., Klawonn, F., Kruse, R., and Runkler, T. (1999). Fuzzy Cluster Analysis: Methods
for Classification, Data Analysis and Image Recognition, John Wiley, West Sussex.
Jarvis, R. A. (1973). On the identification of the convex hull of a finite set of points in the plane.
Info. Proc. Letters. 2: 18–21.
Jefferis E. S., and Mamalian C. A. (1998). Crime Mapping Research Center’s Hot Spot Project.
The Second Annual Crime Mapping Research Conference, December 1998. Arlington, VA.
Kaufman, L., and Rousseeuw, P. (1990). Finding Groups in Data: An Introduction to Cluster
Analysis, John Wiley, New York.
Lawson, A. B. (2001). Statistical Methods in Spatial Epidemiology, John Wiley and Sons,
Chichester.
Levine N. (1999). CrimeStat: A Spatial Statistics Program for the Analysis of Crime Incident
Locations, version 1.0. Ned Levine and Associates/National Institute of Justice, Washington
DC.
Levine N. (2001). CrimeStat: A Spatial Statistics Program for the Analysis of Crime Incident
Locations, version 2.0. Ned Levine and Associates/National Institute of Justice, Washington
DC.
Liu Z. J., and George R. (2003). Fuzzy cluster analysis of spatio-temporal data. Computer and
Information Sciences – ISCIS 2003 – Lecture Notes in Computer Science. 2869: 984–991.
MacQueen J. (1967). Some methods for classification and analysis of multivariate observations.
In Le Cam L. and Neyman, J. (eds.), Proceedings of the Fifth Berkeley Symposium on
Mathematical Statistics and Probability, Vol. I. University of California Press, Berkeley.
Milligan, G. W., and Mahajan, V. (1980). A note on procedures for testing the quality of a
clustering of a set of objects. Decision Sci. 11: 669–677.
Milligan, G. W., and Cooper, M. C. (1985). An examination of procedures for determining the
number of clusters in a dataset. Psychometrika 50(2), 159–179.
Murray, A. T. (1999). Spatial analysis using clustering methods: Evaluating central point and
median approaches. J. Geogr. Syst. 1: 367–383.
Murray, A. T. (2000). Spatial characteristics and comparisons of interaction and median
clustering models. Geogr. Anal. 32: 1–19.
Murray, A. T., and Estivill-Castro, V. (1998). Cluster Discovery Techniques for Exploratory
Spatial Data Analysis. Int. J. Geogr. Inf. Sci. 12: 431–443.
Murray, A. T., and Grubesic, T. H. (2002). Identifying non-hierarchical spatial clusters. Int. J.
Ind. Eng. Theory Appl. Practice 9(1), 86–95.
Openshaw, S. (1984). The modifiable areal unit problem. Concepts and Techniques in Modern
Geography, Vol. 38, Norwick, Geo Books.
Preparata, F. R., and Shamos, M. I. (1985). Computational Geometry: An Introduction,
Springer-Verlag, New York.
Ratcliffe, J. H., and McCullagh, M. J. (1999). Hotbeds of crime and the search for spatial
accuracy. Geogr. Syst. 1(4), 385–398.
Ratcliffe, J. H., and McCullagh, M. J. (2001). Chasing ghosts: Police perception of high-crime
areas. Br. J. Criminol. 41: 330–341.
Rousseeuw, P. (1987). Silhouettes: A graphical aid to the interpretation and validation of
cluster analysis. J. Comput. Appl. Math. 20: 53–65.
Rousseeuw, P., and Leroy, A. (1987). Robust Regression and Outlier Detection, John Wiley,
New York.
Unwin, D. J. (1996). GIS, Spatial Analysis and Spatial statistics. Progr. Hum. Geogr. 20(4),
540–551.
Application of Fuzzy Clustering for Crime Hot Spot Detection 105
Vinod, H. (1969). Integer programming and the theory of grouping. J. Am. Stat. Assoc. 64: 506–
517.
Zhang, J. Q. (2004). Risk assessment of drought disaster in the maize-growing region of
Songliao Plain, China. Agric. Ecosyst. Environ. 102(2), 133–153.