International Journal of Scientific Research & Engineering Trends

Volume 2, Issue 5, Sept.-2016, ISSN (Online): 2395-566X

DBSCAN: An Assessment of Density Based Clustering

and It’s Approaches
Karuna Kant Tiwari Virendra Raguvanshi Anurag Jain
Radha Raman Institute of Technology & Radha Raman Institute of Technology & Radha Raman Institute of Technology &
Science, Bhopal, India Science, Bhopal, India Science, Bhopal, India
Email: [email protected] Email: [email protected] Email: [email protected]

Abstract - Density based clustering is an emerging field of data mining now a days. There is a need to enhance Research
based on clustering approach of data mining. There are number of approaches has been proposed by various author.
VDBSCAN, FDBSCAN, DD_DBSCAN, and IDBSCAN are the popular methodology. These approaches are use to ignore the
information regarding attributes of an objects. This paper is collection of various information of density based clustering. It
also throws some light on the DBSCAN.

Keywords – Data Mining, Clustering, DBSCAN.

I. INTRODUCTION "World Understanding our requires the conceptualization

of similarities and differences between the entities that
Today, data is automatically received from various types compose.
of equipment. Satellites, X-rays and traffic cameras are
some of them. For the information / data understandable to III. CLASSIFICATION OF CLUSTERING
us, should be treated. When working with large data sets is ALGORITHM
useful in many scenarios to separate the information by
dividing the data into smaller categories, and finally the There are some good clustering algorithms used there;
identification of class. No less important is it important in one of them is the famous CLARANS. Other methods
the treatment of large spatial databases. A satellite, for include K-means, K-medoid, hierarchical clustering and
example, collects an image that moves around our land. self-organizing maps. However, none of these algorithms
You want to classify portions of images of houses, cars, can handle all three problems mentioned in the right
roads, lakes, forests, etc. From the database of the image is direction. This report does not deal with these methods,
large, a good classification algorithm is necessary. but focus on DBSCAN (based spatial clustering
Classification can, for example, is made by means of applications with noise density) [1] algorithm, which
clustering algorithms which similar data is grouped in offers solutions to these problems.
different groups. However, the use of clustering
algorithms involves some problems: It is often difficult to
know which are the input parameters to be used for a
specific database, if the user does not have sufficient
domain knowledge. In addition, spatial data sets can Clustering Approach
contain large amounts of data, and try to find patterns of
the various cluster sizes is very computationally
expensive. Short calculation time is always favorable.
Finally, the shapes of the groups may be arbitrary, and in
severe cases very complex. Find these forms can be very
heavy.
Partitioning Hierarchy
II. CLUSTERING ALGORITHM
Clustering and classification are two fundamental tasks
of data mining. The classification is primarily used as a Fig.1. Basic Classification of Clustering Approach
supervised learning method, the combination of
unsupervised learning (clustering models are both). The Partitioning Algorithm:
goal of the group is descriptive classification is predictive. Construct various partitions then evaluate them by some
Since the objective of the group is to find a new set of criterion (CLARANS, O(n) calls). This type of algorithm
categories, new groups are of interest in themselves, and constructs a partition of a database D of n objects into a set
evaluation is intrinsic. In classification tasks, however, a of k groups. k is an input parameter for these algorithms is
large part of the assessment is extrinsic, because the that the domain knowledge is unfortunately not available
groups must reflect a certain set of reference classes. for many applications is needed. The partitioning

© 2016 IJSRET
109
 International Journal of Scientific Research & Engineering Trends
Volume 2, Issue 5, Sept.-2016, ISSN (Online): 2395-566X

algorithm usually starts with initial partition D and then number of objects required for a cluster, it is marked as
uses an iterative control to optimize an objective function. core object and if the objects in it surrounding within
Each group is represented by the center of gravity of the given Eps are less than the minimum number of objects
cluster or a group of objects located near its center. required, then this object is marked as noise. The search
Accordingly, the separation algorithms using a procedure continues for all the objects in the dataset. Later on if the
in two stages. First, determine the k representatives minimum numbers of objects within a given radius are met
minimizing the objective function. Second, assign each subsequently previously marked objects as noise are
object in the class with his "closest" representative of the renamed, in this way the DBSCAN differentiate between
object in question. The second step involves a partition is the border points of a cluster and noisy objects.
equivalent to a Voronoi diagram and each group is
contained in one of the Voronoi cells. Therefore, the form V. THE DBSCAN ALGORITHM
found in all groups by a partitioning algorithm is very
restrictive convex. The DBSCAN algorithm can identify clusters of large
Hierarchy Algorithm: spatial data sets watching the local density of blocks of data
Create a hierarchical decomposition of the set of data (or using a single input parameter. In addition, the user gets a
objects) using some criterion (merge & divisive, difficult suggestion that the parameter value which would be
to find termination condition).In the hierarchical appropriate. Therefore, a minimum area of knowledge is
decomposition of D. The hierarchical decomposition is required. The DBSCAN can also determine what
represented by a dendrogram, a tree that is iteratively information should be classified as noise or outliers.
divided into smaller subsets until each subset D of a single Despite this, it is the work process is fast and scales well
object is made. In such a hierarchy, each node of the tree with the size of the database almost linearly. By using the
represents a group of D. The dendrogram can be created density distribution of nodes in the database, those nodes
from the leaves to the root (agglomeration approach) or DBSCAN be classified into distinct groups defining
from the root to the leaves (approach of division) merger different classes. DBSCAN can find clusters of arbitrary
or division of groups in each step. Unlike separation shape, as shown in Figure 1 [1]. However, groups that are
algorithms, algorithms need not hierarchical k as input. close together tend to belong to the same class.
However, a condition of termination must be set indicating
that the process of merger or division must be completed.
An example of a termination approach agglomeration state
Dmin is the critical distance between all groups Q. Until
now, the main problem with hierarchical clustering
algorithms has been the difficulty of deriving the
appropriate settings for the termination condition for
example, a value of Dmin is small enough to remove all
the "natural" groups, while large enough so that no group
is divided into two parts. Recently, in the field of signal
processing Ejcluster hierarchical algorithm was presented
automatically derive a termination condition. Its main idea
is that two points belong to the same group if you walk in
the first point of the second stage of a "sufficiently small".
Ejcluster follows the approach of the division. It requires
no intervention by domain knowledge. In addition,
experiments show that is very effective in the discovery of
non-convex groups. However, the computational cost of
Ejcluster is O (n2) due to the calculation of the distance
for each pair of points. This is acceptable for applications
such as character recognition with moderate values of n,
but is prohibitive for applications in large databases.

IV. DBSCAN Fig.2. Example of Density based Clustering

DBSCAN (Density based spatial clustering of

V. APPLICATIONS OF DBSCAN
application with noise) [14] is density based method which
can identify arbitrary shaped clusters where clusters are
An example of software program that has the DBSCAN
defined as dense regions separated by low dense regions.
algorithm implemented is WEKA. The following of this
DBSCAN starts with an arbitrary object in the dataset and
section gives some examples of practical application of the
checks neighbor objects within a given radius (Eps). If the
DBSCAN algorithm.
neighbours with in that Eps are more than the minimum

© 2016 IJSRET
110
 International Journal of Scientific Research & Engineering Trends
Volume 2, Issue 5, Sept.-2016, ISSN (Online): 2395-566X

Satellites Images: quantized in finished cells which form the network

A large number of satellite data is received worldwide structure and perform the grouping number of networks.
and these data must be translated into understandable On the basis of the cluster network assigns infinite number
information, eg, classification of satellite images taken in of data records in the data flow to a finite number of
accordance areas with forests, water and mountains. Before networks. In this article, the clustering algorithms using
the DBSCAN algorithm can classify these three elements algorithms based on a grid based on the concept of the
in the database, a work must be done with image density or density are considered for clustering. We call
processing. Once the image processing is given, the data grouping algorithms network density. We explore
appears as spatial data where DBSCAN can sort the groups algorithms in detail and the advantages and limitations of
if desired. them. The algorithms are also summarized in a table on the
X-ray crystallography: basis of important features. In addition, it describes how
X-ray crystallography is another practical application well the algorithms deal with difficult issues in clustering
that locates all the atoms in a crystal, which causes a large data streams.
amount of data. The DBSCAN algorithm can be used to In this article, based on the density outlier mining
find and classify the atoms in the data. similarity-neighbor technique for data preprocessing for
Anomaly Detection in Temperature: data mining algorithm is proposed. First, the notion of a k-
This type of application data focuses on the anomalies in density object is presented, and the number of similar
the data model, which is important in many cases, for density (SDS) of the object to the evolution of the density
example, credit fraud, health, etc. This application of objects and densities is established on the neighboring
measures the temperature anomalies [3], which is base. Second, the series of average cost (AUC) of the
important because of environmental changes (global object on the basis of the weighted sum of the distance
warming). You can also find computer errors and so on. between adjacent objects to the object is obtained SDS.
These unusual patterns must be identified and examined to Finally, the density of outliers similarity factor based on
take control of the situation. The DBSCAN algorithm is neighbor (DSNOF) of the object is calculated using both
able to discover these patterns in the data the AUC and the object of the k-ASC neighbor distance
object, and the degree of object is an outlier is indicated by
VI. LITERATURE REVIEW DSNOF. The experiments were performed on sets of
synthetic data and real data to evaluate the effectiveness
Classification method of research based on the density is and performance of the proposed algorithm. The results of
an important task of data mining. To improve methods experiments to demonstrate that the proposed algorithm has
based on the density of the space attribute (such as an outlier extraction better and do not increase the
DBSCAN, Camarilla, optical, etc.) which do not take into complexity of the algorithm.
account the relationship between objects and methods Social networking has been considered a timely and cost
based on the density of the network (eg , SCAN, DCSBRD effective source of spatio-temporal information for many
etc.) to ignore the attribute information of the object, a fields of application. However, while some research groups
clustering algorithm based on density weighted network have successfully developed methods for detecting current
attribute information (WN DCA) proposed in the issue of the text for a while, and even some popular
document. After setting the network-based distance microblogging services such as Twitter to provide
weighted attribute, the algorithm updates the definition of information of the main trend themes for the selection
the object nearest neighbor and the object, and provides the remains incapable of fully support users to collect all the
appropriate policy group. To take into account both the items on the events in real time with a point of complete
information of attributes and relations, the algorithm spatio-temporal view to meet their information needs. This
increases the accuracy of clustering improves the results of work aims to study how micro-blogging social network (eg
clustering, and distinguish the nature and aberrant objects Twitter) can be used as a reliable source of emerging
effectively. events to extract the spatio-temporal characteristics of
The data stream clustering attracted many researchers messages to increase awareness event information. In this
and applications that generate data streams have become paper, a method for online classification based on density
more popular. Several clustering algorithms have been flow micro blogging text mining, in order to obtain spatial
introduced for data streams based on the distance they are and temporal characteristics of real-world events are
incompetent to find clusters of arbitrary shapes and can not applied. By analyzing events detected by our system,
handle outliers. Classification algorithms based on density temporal and spatial impacts of emerging events can be
are remarkable not only for finding clusters arbitrarily, but estimated for the attainment of situational awareness and
also to deal with noise in the data. In classification risk management.
algorithms based on the density of the dense zones of the The advent of modern for scientific data collection
objects in the data space are considered groups are techniques has led to the massive accumulation of data
separated by areas of low density. Another group of from various fields. Cluster analysis is one of the main
methods of classification of data streams is based on the methods of data analysis. It is the art of all similar items in
combination of the network wherein the data space is large data sets without the need to detect the specified

© 2016 IJSRET
111
 International Journal of Scientific Research & Engineering Trends
Volume 2, Issue 5, Sept.-2016, ISSN (Online): 2395-566X

groups by explicit functions. The problem of detection is VIII. CONCLUSION

difficult when the groups are of different size, density and
shape. This paper provides a new approach to clustering Clustering is most famous approach of data mining in
based on the approach of the density. DBSCAN is order to create new sub classes which are known as the
considered one of the pioneers of density on the technical clusters. In this way it seems to be that the density based
basis of clustering; this paper makes a step towards the clustering can also apply. In this study, we have presented
detection of groups within a cluster. On the basis of various the summary information of the different enhancement of
parameters necessary for proper clustering algorithm is density-based clustering algorithm called the DBSCAN.
estimated that the number of groups formed, the noise in The purpose of these variations is to enhance DBSCAN to
the change of distance, time to form a group where non- get the efficient clustering results from the underlying
cluster and incorrectly. datasets. In addition, we also have highlighted the research
contributions and found out some limitations in different
VII. RESEARCH SCOPE research works. Consequently, this work also depicts the
critical evaluation in which comparison and contrast have
This section primarily reflects the comparison and been taken out to show the similarities and differences
contrast of the above reviewed literature regarding the among different author’s works. The spatiality of this work
different DBSCAN variations and modifications. It is that it reveals the literature review of different DBSCAN
identifies the similarities and differences among the various modification and provides a vast amount of information
research works on the DBSCAN algorithm enhancements. under a single paper. In our future work, we have planned
This will help for the future research in the DBSCAN to enhance the DBSCAN and provide its implementation
modification and enhancements. and compare its results with the different existing
Liu et al. [11] have modified the DBSCAN to deal with DBSCAN algorithms variations.
the datasets that are varied in densities. Their algorithm is
called VDBSCAN. VDBSCAN is able to calculate the ACKNOWLEDGEMENT
density threshold parameters automatically based on the K-
distance plotting. Its computational complexity is same as I would like to say thanks to my guide “Virendra
that of DBSCAN. The same work is explored in Raguvanshi” who gives their knowledge and time in order
GRIDBSCAN [12] to deal with the dataset that have to complete this paper. This paper will never complete
cluster with different densities. The research work without the support faculty member of CSE department of
proposed in [11, 12] are identical in that they do not require college name, Bhopal.
any user supplied input parameters. The study carried out
by [12] can cluster the dataset efficiently as that of [11] but REFERENCES
[12] is expensive as compare to that of [11]. Fahim et al.
[13] carried out the research in the same dimension as that [1] H. Sun, J. Huang, J. Han, H. Deng, P. Zhao, and B. Feng,
of [11] in the sense that it does not require any user “Gskeletonclu: Density-Based Network Clustering via
supplied density threshold parameters. Structure-Connected Tree Division or Agglomeration,”
Uncu et al. [12] have introduced an extension of IEEE 2010, pp. 481-490, 2010.
DBSCAN such that it can cluster the datasets having [2] M. Girvan and M.E.J. Newman, “Community Structure in
different densities. The author in [12] has used the concept Social and Biological Networks,” IEEE 2002, vol. 99, no.
of grid while performing clustering. Its clustering results 12, pp. 7821-7826.
are more efficient than results produced by DBSCAN. [3] A. Clauset, C. Moore, and M.E.J. Newman, “Hierarchical
Structure and the Prediction of Missing Links in
Similar grid based technique is also used by Mahran et al. Networks,” IEEE 2008, vol. 453, pp. 98-101
[14] to generate efficient clustering output from the [4] J.M. Kleinberg, “Authoritative Sources in a Hyperlinked
underlying dataset and it has proved more faster than Environment,” IEEE 1998, pp. 668-677
DBSCAN . The method in [12] was more costly than that [5] P. Domingos and M. Richardson, “Mining the Network
of [14] when applied on the large volume of datasets. Value of Customers,” ACM 2001, pp 57-66
YU et al. [15] also used the local density in its [6] Wu Lingyu, Gao Xuedong, “A Density-based Clustering
clustering technique for large datasets. EDBSCAN [16] Algorithm for Weighted Network with Attribute
also focused on the local density variation and provided an Information”, 3rd International Conference on Advanced
enhancement to DBSCAN. Computer Control, IEEE 2011, pp 629-633.
[7] Amineh Amini, Teh Ying Wah, Mahmoud Reza Saybani,
The clustering techniques described in [14, 15] have Saeed Reza Aghabozorgi Sahaf Yazdi, “A Study of
achieved the efficient clustering result by using the local Density-Grid based Clustering Algorithms on Data
density in their clustering technique. The density-based Streams”, Eighth International Conference on Fuzzy
techniques discussed in [11, 12] does not need density Systems and Knowledge Discovery, IEEE 2011, pp 1652-
threshold to be input by the end users. The technique 1656.
described in [16] requires the user input density threshold [8] Hui Cao, Gangquan Si, Yanbin Zhang and Lixin Jia,
manually. “Enhancing effectiveness of density-based outlier mining
scheme with density-similarity-neighbor-based outlier
factor”, Expert Systems with Applications, elsevier 2010,

© 2016 IJSRET
112
 International Journal of Scientific Research & Engineering Trends
Volume 2, Issue 5, Sept.-2016, ISSN (Online): 2395-566X

pp 8090-8101
[9] Chung-Hong Lee, “Mining spatio-temporal information
on microblogging streams using a density-based online
clustering method”, Expert Systems with Applications,
elsevier 2012, pp 9623–9641
[10] Glory H.Shah, “An Improved DBSCAN, A Density Based
Clustering Algorithm with Parameter Selection for High
Dimensional Data Sets”, IEEE 2012,pp 1-6.
[11] P. Liu, D. Zhou, and N. J. Wu,“VDBSCAN: Varied
Density Based Spatial Clustering of Applications with
Noise,” in proceedings of IEEE International Conference
on Service Systems and Service Management, Chengdu,
China, pp 1-4, 2007.
[12] O. Uncu, W. A. Gruver, D. B. Kotak, D. Sabaz, Z.
Alibhai, and C. Ng, “GRIDBSCAN: GRId Density-Based
Spatial Clustering of Applications with Noise,” 2006
IEEE International Conference on Systems, Man, and
Cybernetics October 8-11, 2006, Taipei, Taiwan.
[13] A. M. Fahim, A. M. Salem, F. A. Torkey, and M.A.
Ramadan, ”Density Clustering Based on Radius of Data
(DCBRD),” World Academy of Science, Engineering and
Technology 2006.
[14] S. Mahran and K. Mahar, “Using Grid for Accelerating
Density Based Clustering,” Computer and Information
Technology, CIT2008, 8th IEEE International Conference
on. 08/08/2008, ISBN: 978-1-4244-2357-6, Sydney,
NSW.
[15] X. P. Yu, D. Zhou, and Y. Zhou, “A New Clustering
Algorithm Based on Distance and Density,” presented in
proceedings of International Conference on Services
Systems and Services Management (ICSSSM-2005), Vol.
2.
[16] A. Ram, A. Sharma, A. S. Jalall, R. Singh, and A.
Agrawal, “An Enhanced Density Based Spatial Clustering
of Applications with Noise,” 2009 IEEE International
Advance Computing Conference (IACC2009) Patiala,
India, 6-7 March 2009.

© 2016 IJSRET
113