Spatial Clustering Algorithm Using R-Tree
Spatial Clustering Algorithm Using R-Tree
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING
WWW.JOURNALOFCOMPUTING.ORG 85
Abstract — Recently, there has been extensive growth in databases; we require proper analyzing techniques to dis-
cover hidden patterns from large databases. Spatial data mining is the application area of data minining which helps to
detect hidden patterns from spatial datasets. There has been enormous growth in spatial data from last decade, growing
need has forced researchers to find efficient algorithm for spatial data analysis. There are several spatial indexing tech-
niques applied in past for storage and retrieval of spatial data. In this approach suitable indexing techniques has being
proposed for three dimensional spatial objects such as R tree which has been used as a spatial data structure for index-
ing of spatial data. We have designed a spatial clustering model using R tree (SCART), with three dimensional spaces
using spatial objects where data is organized in the form of three dimensional grid and Hilbert space filling curve is used
to find the linear order points of the grid. In this paper efficient clustering algorithm with combination of hierarchical and
grid based approach has being used to find the effective and efficient spatial cluster. The structure of node has been de-
veloped to find the relevant clusters with the help of R tree data structure. The novel search algorithm has being pro-
posed for finding the objects or retrieve the answer for given query. We have found that the algorithm is suitable for
achieving three dimensional searching in a grid.
Index Terms— Spatial Databases, Three Dimensional Grid, R -tree, Spatial Cluster Algorithm using R tree (SCART),
Space filling curve
—————————— ——————————
1 INTRODUCTION
collected from NASA satellites, Geographical Information
K DD (Knowledge Discovery in Databases) is the pro‐
cess to extract knowledge from data, whereas data
mining is the step in KDD process to determine pat‐
System, data collected from X‐ray and weather and cli‐
mate data. A spatial database can be called as collection of
spatial attribute and non spatial attribute. Spatial attribute
terns from databases. KDD is a multistep process which includes the location based attributes whereas, Non spa‐
requires data selection, data cleaning, Pre‐processing of tial attribute are age, sex and marital status. There exist
data, data transformation, data mining and interpretation several spatial data mining techniques such as classifica‐
of results in mining. Data mining can be defined as the tion, visualization, clustering and spatial association. The
non‐trivial process of identifying valid, novel, potentially aim of spatial data mining is to achieve the best
useful and ultimately understandable patterns in data knowledge of spatial data and retrieve the hidden infor‐
and describing them in a concise and meaningful way. mation from spatial databases. The features in spatial da‐
Data mining involves interest for researchers in fields ta are represented in the form of point, line and polygon.
such as statistics, artificial intelligence, machine learning, Spatial databases can be stored in 2‐dimensional or N‐
data visualization techniques, classification and pattern dimensional spaces. To deal with several dimensionalities
recognition [12], [13], [14], [21]. of spatial data numerous spatial data structures has being
Researchers are faced with challenges to handle large proposed in the past [27]. The contribution of our work
amount of spatial data and relative algorithm which can focuses on R‐tree as spatial data structure to discover spa‐
effectively and efficiently scale the data. To deal with tial clusters. An R‐tree is spatial data structure which does
enormous amount of spatial data, we have spatial data not require transforming point data and hence provide
mining technique. Spatial data mining refers to discovery better feasibility for spatial clustering algorithms. It has
of knowledge from large spatial databases. Examples, of attracted researchers from past deacde for its robustness
spatial databases include remotely sensed images, data to handle large amount of multidimensional spatial data
and its efficiency has proved beneficial for several query
————————————————
algorithms. There is certain disadvantage of R tree as it
Harleen Kaur is with the Department of Computer Science, Hamdard Uni-
lacks from poor updation problem. To overcome this
versity, New Delhi, India. flaws R*‐tree [9] was discovered, it works similarly like R
Ritu Chauhan is a Research Scholar with the Department of Computer tree data structure, but it has updation algorithm that
Science, Hamdard University, New Delhi. certainly makes it more efficient than R trees.
M. Afshar Alam is with the Department of Computer Science, Hamdard
University, New Delhi, India. The rest of the paper is organized as follows. Section
2 briefly gives background knowledge explaining drift in
spatial data mining and several spatial data structures for
JOURNAL OF COMPUTING, VOLUME 3, ISSUE 2, FEBRUARY 2011, ISSN 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING
WWW.JOURNALOFCOMPUTING.ORG 86
clustering of spatial data. In Section 3, we describe several ally separated by regions of low density which can also
definitions based on spatial clustering algorithm as well contain noise. Density based approach has being exten‐
as R tree as index structure in three dimensional spatial sively applied to deal with spatial databases. The density
grids. It explains spatial cluster algorithm using R tree based clusters are generally formed with set of objects
(SCART) model and structure for three dimensional R that meets some density requirement. They are several
trees. Section 4 explains the role of space filling curve in pre existing density based clustering algorithms approach
grid data structure and define search algorithm in three which relates with spatial data such as Density‐Based
dimension space and the challenges encountered, Section Spatial Clustering of Applications with Noise (DBSCAN)
5 conclusion is covered and in the last section future work [7] and Distribution Based Clustering of Large Spatial
is presented. Databases (DBCLASD) [25]. The grid based clustering
algorithm divides the region into finite space and clusters
are discovered according to adjancancy of neighbour.
2 LITERATURE REVIEW There are several grid based clustering algorithm such as
There has been tremendous increase in the spatial data Statistical Information Grid (STING) [22] which is a multi
for the last two decades; researchers are laying down resolution grid based clustering algorithm, it divides the
enormous efforts to detect hidden patterns from spatial region space into finite number of cells and spatial clus-
databases [6]. The outcome of research has been used in ters are retrieved using statistical techniques such as
numerous application areas such as medical domain, Clustering in Quest (CLIQUE) is combination of density
business analysis, weather forecasting, etc. Spatial Data and grid based approach to find clusters in high dimen-
Mining can be used to discover unknown hidden patterns sionality of spatial data [2], Wave Cluster algorithm [20]
from spatial databases. There are several pre existing clusters data in multidimensional grid space and wavelet
techniques used in discovery of spatial patterns such as transform determines the dense regions of space and
spatial clustering, spatial association rules, spatial charac‐ STING+ [24] is enhanced part of STING algorithm. In this
terization and spatial trend detection [14]. approach triggers are used whenever there is change in
Spatial clustering plays major role in spatial data min‐ the database. Spatial queries are processed with the pyr‐
ing techniques. It has being widespreadly used for several amid like structure. The main advantages of grid based
application domains such as weather forecasting, remote clustering it has fast processing time, generally depends
sensing, satellite image analysis and several other re‐ on the dataset. Spatial data requires proper indexing for
search areas. Spatial clustering can be defined as unsu‐ storage, updation and querying of data. To deal with such
pervised learning technique because it does not work on key issues several spatial data structures has being pro-
criteria of predefined class. Spatial data clustering is the posed in past to deal with large spatial datasets.
process of grouping the objects according to the similarity The Spatial data structures are of significant im‐
of spatial objects. The objects inside the cluster have high portance for storange and retrieval of spatial data for sev‐
similarity whereas outside the cluster objects are highly eral application domains [15]. They are several pre exist‐
dissimilar in their properties [1]. The spatial clustering ing spatial data structures such as quad tree utilized for
algorithms are capable of discovering clusters of variant two dimensional spaces [26] it can be defined as the hier‐
shapes. The clustering algorithms can be broadly catego‐ archical spatial data structure which divides the space
rized into Partitioning, Based, Hierarchical based, Density into four equal sized nodes. The quad tree consists of
Based and Grid based clustering [8],[ 6]. They are recently unique root where each parent nodes divides its self re‐
reviewed clustering techniques can be found in [2], [6], cursively into four child nodes whereas R tree is a height
[4], [14]. balanced tree which is an index based data structure,
Spatial clustering algorithms are designed to deal widely used in application areas of spatial Database for
with large amount spatial data set to find relavant pat- decision making process [3]. There are several different
terns. They are several advances in spatial clustering al- subclasses of R tree which are used for indexing such as
gorithm from the past decade. The first spatial clustering R*Tree [9], Packed R Tree, Hilbert R ‐tree [5] and R+ Tree.
technique developed in spatial data mining to discover K‐d‐tree is a spatial data structure which divides the re‐
patterns from spatial data sets was CLARANS. It can be gional space into two hyperrectangles where the division
defined as Clustering Large Applications based upon of plane is conducted by using perpendicular axis for di‐
Randomized Search (CLARANS) as in [17], [18], [19]. viding the hypeplane through median point [23]. Octree
CLARANS detect patterns based on randomized search has being promisingly used for indexing of spatial data
of data and can easily and efficiently handle large spatial [16] it subdivides the regional space into octant, until or
data. They are several other extensions of CLARANS al- unless cubes are obtained, it’s a tree data structure in
gorithms such as Spatial Dominant Approach SD- three dimensional spaces. In this paper we are focusing
CLARANS and Non-Spatial Dominant NSD CLARANS on R tree as spatial data structure for storage of spatial
which partitions the databases on the basis of K-Medoids data as R tree does not divide space into several pieces as
partition method. There are several density based ap‐ other spatial data structure such as quad tree. We have
proaches applied to spatial data to discover clusters on used several statistical techniques for spatial clustering of
the concept of density for specific region. Spatial clusters data in three dimensional spaces using R tree as spatial
are determined on the basis of dense regions and are usu‐ data structure.
JOURNAL OF COMPUTING, VOLUME 3, ISSUE 2, FEBRUARY 2011, ISSN 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING
WWW.JOURNALOFCOMPUTING.ORG 87
level is determined with 30, 31, 32, until the required level
this can be done, when all the nodes contain maximum
number of entries. The root contains 1 node, level two
3 SPATIAL CLUSTER ALGORITHM USING R TREE IN contains 6 nodes. The number of level depends on the
THREE DIMENSIONAL GRID STRUCTURE required size of the data. We used three dimensional grids
to store the values of dataset. M is the maximum number
Growing data and dimensions has created researchers to
of nodes and m is the minimum number of nodes. If the
adopt new methodology for spatial data analysis. There
number of nodes inserted is less than m then the condi-
are several research based methodologies on two dimen-
tion of underflow exists whereas if it exceeds the M then
sional data, but real time datasets are multidimensional.
overflow of stack occurs. The underflow occurs during
Our proposed work discusses the issues related to spatial
the deletion process whereas overflow occurs during the
data in three dimensional spaces. The proposed algorithm
insertion process.
is based on simple stastical techniques to determine spa-
tial clusters using R tree as spatial data structure. The idea
comes from hierarchical as well as grid based clustering
algorithm. There are several definitions based on algo- Level 0 Root Node
Parent
Node
In three dimensional spaces each data has been regarded Child
Node
as point in spatial grid where each point in N dimensional
Level 3
a) The dataset M in the N dimensional space can be di‐
Fig.1. Three Way R-trees showing different level of distribution
rected as M {m1, m2. m3……, mn} with N number of di‐
mensionality as N {n1, n2, n3}.
Fig. 2 represents the X, Y and Z plane in three dimen-
b) The mean centre of each dimension can be defined as
sional. The grid size can be N*N*N where N represents
summation of n1 {na+ nb…..+ nn}/number of elements
integer, the number of cells in each grid is N3. The over‐
whereas the mean centre of each dimension can be statis‐
lapping of MBR can be represented in three dimensional
tically similarly found.
grids.
c) The dispersion of each cluster formed around mean
centre is calculated as standard dispersion in three di‐
mensional spaces, dis, and relative threshold has been
regarded as between the dimensionality
as 0 . 5 1 , the value is defined between standard
Z plane
Enclosed
MBR
dispersion of value 0.5 and +1. X plane
Y plane
d) The neighbor hood of each grid value is compared and
hence merged if, standard dispersion around mean centre
is substantially between the required threshold values.
The spatial data structure R-tree in three dimensional Fig.2. Enclosed MBR in Three dimensional grid
spaces to represent spatial data in grid structure. Each cell
in grid structure corresponds to spatial attribute value. 3.2 Structure of Spatial Cluster Algorithm using R
The values in these cells are then analyzed with the help tree (SCART)
of R trees and searching is performed. R Tree is a height
balanced tree with leaf level in the tree corresponds to the The operation of R‐tree in three dimensional grid struc‐
actual data present in the spatial grid and the root node tures is stored in main memory. We have partitioned the
summarizes the data in grid. Each node in the R tree di- space of cells in grid according to one diameter in space.
vides itself into maximum of three corresponding nodes, Grid helps to find adjacency points and neighbors very
except the root level corresponds to minimum of two easily. The leaf nodes in tree point to the spatial objects.
nodes, until or unless it is a leaf node. The splitting of The Spatial cluster algorithm using R tree (SCART) merg‐
node occurs when each data points occur and it organizes es the cluster as similar to agglomerative hierarchical
itself on the grid structure. A leaf node contains spatial cluster but has spatial constraints related to spatial ob‐
object as well as MBR that is the minimum bound rectan- jects. The spatial objects are merged together as we go
gle of spatial object. A non leaf has list of child node and through bottom up approach or reverase approach,
minimum bound rectangle. The bounding rectangle in whereas the R tree is formed in forward approach which
parent node contains all the rectangle boxes of child node. is from top to bottom. The parent node contains the in‐
Fig. 1 represents R tree data structure where numbers of formation of child node as we move in forward approach
nodes are determined at each level. Each node divides from one level to another. The clusters are formed on the
itself to three nodes at each level until all the data points basis of the information processed on the grid structure
are located in the grid form. The number of nodes at each from reverse approach. The R tree structure connects eve‐
JOURNAL OF COMPUTING, VOLUME 3, ISSUE 2, FEBRUARY 2011, ISSN 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING
WWW.JOURNALOFCOMPUTING.ORG 88
ry point on the grid at various levels to form relevant clusters at different level of tree. The clusters of common
clusters. We recursively keep on expanding the tree until characteristics are prepared at each node at various levels.
or unless the spatial objects are sufficiently placed on grid We start our algorithm after we have calculated statistical
structure. The inner node in R tree contains the statistical relevant factors at each node which mainly includes mean
information such as mean centre and standard dispersion centre value, child nodes, parent node, level, standard
of all the child nodes. This helps us in retaining the rele‐ dispersion and x, y, and z coordinates. Each node consists
vant clusters and discards the others. of list of values and the relevant lists are found and merge
The structure of R tree node in three dimensional archi‐ from one level to other in bottom up fashion. The algo‐
tecture of spatial grid is defined as rithm to calculate the relevant information in bottom up
approach is as follows:
Struct node {
TypeMBR MBR, // minimum bound rectangle Algorithm Find_Clusters
Int node_number, // node number of each node If (node_level >= 0)
Float xcoordinate, // value of xcoordinate 1. For each node calculate statistical relevant factor
Float ycoordinate, // value of ycoordinate 2. mean centre
Float zcoordiante, //value of zcoordinate 3. standard dispersion , the number of child node ,
Bool lnode, //determine the node is leaf or non leaf
the x, y, z coordinate values and minimum
Float mean_centre, //determine the mean centre of
bounding rectangle
node
Float standard dispersion// measure of dispersion 4. merge the list according to threshold value
Int listofitems, // total number of elements in node 5. else
Int parent_number, //the parent node 6. { increase the node_level and then repeat the
} find_cluster}
7. End
The statistical relavant information is calculated at each
level by SCART algorithm to determine the correspond‐
ing leaf nodes. We determine the mean centre as well as 4 SPACE FILLING CURVE
the standard dispersion of each node at each level of tree.
The mean centre contains the attribute values of nodes at The grid based approach is very efficient to handle large
lowest level, the mean centre has the least square proper‐ amount of databases. It utilizes the space into finite num‐
ty which can be defined for three dimensional spaces by ber of blocks and it focuses on large number of blocks. To
(1), (2) and (3) where X m is the mean centre represents arrange the data on the grid we are using space filling
X i the attribute values and n is the total number of val‐ curves. The space filling curves are the continuous curves
ues, correspondingly we are calculating the mean centre which moves exactly once among the points in grid with‐
of other dimensions. out crossing themselves. The curves usually have two free
ends that may be joined with other paths. We are using
X
Xi
(1) the Hilbert space filling curve to store the data on three
m n dimensional grid structures. It is first described by [10]
and used to express the locality of two‐dimensional data
(2)
Ym
Yi in a one‐dimensional space. The Hilbert space‐filling
curve imposes a linear ordering on the grid cells, so that
n
similar data should be placed together in the linear order
Zm
Zi
(3)
whereas assigning single value to each cell. The Hilbert
space filling curve achieves the best clustering by mini‐
n mizing the number of clusters. The search algorithm im‐
plemented on three dimensional R tree as follows.
The standard dispersion of node n can be represented
as dis , where we are finding the absolute dispersion of 1. The searching algorithm starts with the root node
objects in space. The measure of standard dispersion is
G where the cuboid H is the search and to find
given in (4).
each cuboid overlapped by the search cuboid.
2. Search Sub tree: if the root node is not a leaf node
dis
( X i X m ) 2 (Yi Y m ) 2 ( Z i Z m ) 2
(4) then each entries in the sub tree L, whether it
n ( n 1) overlaps the cuboid H or not. If it overlaps then
we start search for the consecutive non leaf node.
The calculated information is propagated above the R 3. Search leaf node: if the root G is a leaf node, then
tree from bottom to up approach; the similar values are we check all entries for the suitable overlap and
merged as relevant to information. We have found the result is found.
mean centre, standard dispersion of all node values with
the top to bottom approach. We aim to discover similar
JOURNAL OF COMPUTING, VOLUME 3, ISSUE 2, FEBRUARY 2011, ISSN 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING
WWW.JOURNALOFCOMPUTING.ORG 89
5 CONCLUSIONS 4th International Conference on Knowledge Discovery and Data
Mining (KDD98), New York, NY, USA, pp. 58‐65, August 1998.
The approach is to find R tree as a suitable index structure
[12] Jajuga, K., Sokolowski, A., Bock, H. ‐H. Classification, cluster‐
for three dimensional grids. We have demonstrated a ing and data analysis. New York: Springer, 2002.
search algorithm in three dimension grid with the help of [13] Kaur, H., Wasan ,S. K., An Integrated Approach in Medical
R tree as index data structure for Decision Making for Eliciting Knowledge, Web‐based Applica‐
Spatial data, the Hilbert space filling curve is used for tions in Health Care & Biomedicine, Annals of Information Sys‐
three dimensional grids to store the values. The spatial tems (AoIS),U.S.A., Eds: A. Lazakidou, Springer, 2009.
clustering algorithm is developed to find the clusters of [14] Kaur, H., Chauhan, R., Alam, M, A., An Optimal Categorization
different shapes without any user input. The R tree data of Feature Selection Methods for Knowledge Discovery, Visual
structure is really efficient to deal with the spatial data Analytics and Interactive Technologies: Data, Text and Web
and hence can prove to be backbone for future spatial Mining Applications, Eds: Zhang, Segall, and Cao, IGI Publish‐
data mining techniques. ers Inc., 2010.
[15] Samet, H., The Design and Analysis of Spatial Data Structures.
Addison–Wesley, 1990.
6 FUTURE WORKS [16] Fujimura, K., Toriya, H., Tamaguchi, K.., Kunii. T. L., Octree
The future work will be focused on other spatial index algorithms for solid modeling. In Proc. Inter graphics’83, vol‐
techniques that can be helpful for determination of spatial ume B2‐1, pages 1–15, 1983.
cluster with the help of spatial data structures and space [17] Kaufman, L., Rousseeuw, P.J., Finding Groups in Data: an In‐
filling curve for n dimensional data space. troduction to Cluster Analysis, John Wiley and Sons, 1990
[18] MacQueen, J., Some Methods for Classification and Analysis of
Multivariate Observations, In Proceedings of 5th Berkeley Sym‐
posium on Mathematical Statistics and Probability, 1, pp. 281‐
REFERENCES
297.
[1] Abonyi, J, Feil, B., Cluster analysis for data mining and system [19] Ng, R. T., Han, J., Efficient and Effective clustering methods for
identification. Boston, MA: Birkhäuser Basel, 2007. spatial data mining, In Proceedings of 20th International Con‐
[2] Agrawal, R., Gehrke, J., Gunopulos, D., Raghavan, P., Automat‐ ference on Very Large Data Bases, Santiago, Chile, pp. 144‐155,
ic Subspace Clustering of High Dimensional Data for Data Min‐ September 12‐15, 1994,
ing Applications, Proceedings of the 1998 ACM‐SIGMOD In‐ [20] Sheikholeslami, G., Chatterjee, S., Zhang, A., WaveCluster: A
ternational Conference on Management of Data, Seattle, Wash‐ Multi‐Resolution Clustering Approach for Very Large Spatial
ington, pp. 94‐105, June 1998. Databases, Proceedings of the 24th Very Large Databases Con‐
[3] Guttman A., R‐trees: A Dynamic Index Structure for Spatial feerence (VLDB98), New York, NY, USA, pp. 428‐439, August
Searching”, Proc. ACM SIGMOD Int.Conf. On Management of 24‐27, 1998.
Data, Boston, MA, pp.47‐57, 1981. [21] ] Hand, D., Mannila, H., Smyth, P., Principles of Data Mining,
[4] Berkhin, P., Survey of clustering data mining techniques. Tech. Prentice Hall of India Private Limited, India, 2001.
rep., Accrue Software, San Jose, CA, 2002. [22] Wang, W., Yang, J., Muntz, R., STING: A Statistical Information
[5] Kamel, I., Faloutsos C., Hilbert R‐tree: An improved R‐tree Grid Approach to Spatial Data Mining, Proceedings of the 23rd
using fractals. In Bocca J., Matthias, J., and Zaniolo C., editors, VLDB Conference, pages 186‐195, Athens, Greece, 1997.
20th International Conference on Very Large Data Bases, Santia‐ [23] Bentley, J. L., Multidimensional binary search trees used for
go, Chile proceedings, pp. 500‐509, Morgan Kaufmann Publish‐ associative searching. Commun. ACM, 18(9):509–517, Sept.1975.
ers , Los Altos, CA 94022, USA, 1994. [24] Wang, W., Yang, J., Muntz, R., STING+: An Approach to Active
[6] Chauhan, R., Kaur, H., Alam, M, A., Data Clustering Method Spatial Data Mining, Proceedings of the Fifteenth International
for discovering clusters in Spatial Cancer Databases, Interna‐ Conference on Data Engineering, Sydney, Australia, pp. 116‐
tional Journal of Computer Applications, 2010. 125, March 23‐26, 1999.
[7] Ester, M., Kriegel, H. –P., Sander, J., Xu, X., A Density‐Based [25] Xu, X., Ester, M., Kriegel, H‐P., Sander, J., A Distribution‐Based
Algorithm for Discovering Clusters in Large Spatial Databases Clustering Algorithm for Mining in Large Spatial Databases,
with Noise, Proceedings of 2nd International Conference on Proceedings of 14th International Conference on Data Engineer‐
Knowledge Discovery and Data Mining (KDD‐96), Portland, ing (ICDE’98), Orlando, FL, pp. 324‐331, AAAI Press, 1998.
Oregon, August 1996, AAAI Press, 1996. [26] Orenstein, J. A., Multidimensional tries used for associative
[8] Everitt, B. S., Landau, S., Leese, M. Cluster analysis (4th ed.). searching. Inform. Process. Lett., 13:150–157, 1982.
London: Arnold, 2001. [27] Samet, H., The Design and Analysis of Spatial Data Structures.
[9] Beckmann N., Kriegel H.P., Schneider R., Seeger B., “The R*‐ Reading, MA: Addison‐Wesley, 1989b.
tree: an efficient and robust access method for points and rec‐
tangles”, Proceedings of the 1990 ACM SIGMOD international Harleen Kaur gained her Ph.D. in Computer Science from Jamia
conference on Management of data, Atlantic City, NJ USA, Millia Islamia University, New Delhi, India on the topic of Applications
of Data Mining techniques in Health care Management. She gradu-
pp.322‐331, 1990.
ated from the University of Delhi, New Delhi. She has previously
[10] Hilbert, D., Uber die stetige abbildung einer linie auf served as a Lecturer in Computer Science, University of Delhi. Cur-
Fl¨achenst¨uck, Mathematische Annalen, vol. 38, pp. 459–460, rently, she is an Assistant Professor at the Department of Computer
1891. Science, Hamdard University. She has published numerous re-
search articles in refereed international journals and conference
[11] Hinneburg, A., Keim, A.D., An Efficient Approach to Clustering
proceedings and chapters in an edited book. She is a member of
in Large Multimedia Databases with Noise, Proceedings of the several international bodies. Her main research interests are in the
JOURNAL OF COMPUTING, VOLUME 3, ISSUE 2, FEBRUARY 2011, ISSN 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING
WWW.JOURNALOFCOMPUTING.ORG 90
fields of Data analysis with applications to medical databases, Medi-
cal decision making, Fuzzy logic, Information Retrieval, Bayesian
networks and visualization.