An Improvement of DBSCAN Algorithm to Analyze

Cluster for Large Datasets

Chetan Dharni Meenakshi Bnasal

Department of Computer Engineering Department of Computer Engineering
Yadavindra College of Engineering Yadavindra College of Engineering
Talwandi Sabo, Bathinda, India Talwandi Sabo, Bathinda, India
[email protected] [email protected]

Abstract— Clustering is an important tool which has seen an problem taking place in huge amount of data in large dataset,
explosive growth in Machine Learning Algorithms. DBSCAN these algorithms are very useful. Clustering for large amount
(Density-Based Spatial Clustering of Applications with Noise) of data is an active research topic for last several years and
clustering algorithm is one of the most primary methods for still is.
clustering in data mining. DBSCAN has ability to find the
DBSCAN (Density Based Spatial Clustering of
clusters of variable sizes and shapes and it will also detect the
noise. The two important parameters Epsilon (Eps) and Applications with Noise) is one of most popular and classical
Minimum point (MinPts) are required to be inputted manually in density based clustering algorithm [5]. DBSCAN algorithm
DBSCAN algorithm and on the basis these parameter the used two important input parameters Epsilon (Eps) and
algorithm is calculated such as number of cluster, un-clustered minimum point (MinPts) and also used no. of cluster, un-
instances as well as incorrectly clustered instances and also clustered instances, incorrectly instances well as time and
evaluate the performance on the basic of parameters selection noise ratio. Density-based clustering algorithms [15, 16] are
and calculate the time taken by the datasets. Experimental proposed based on several concepts including:
evaluation on the basis of different datasets in ARFF format with Core Point: A point is a core point if it has more than
help of WEKA tool which shows that quality of clusters of our
proposed algorithm is efficient in clustering result and more
a specified number of points (MinPts) within Eps. These are
accurate. This improved work on DBSCAN have used in a large points that are at the interior of a cluster.
scope. Border Point: A border point has fewer than MinPts
within specified radius (Eps), but is in the neighborhood of a
Keywords— Machine learning; Clustering; WEKA; DBSCAN; core point.
Noise; Data mining Noise Point: A noise point is any point that is not a
core point or a border point.
I. INTRODUCTION
Other is İ –neighborhood of the point, directly
Data mining is the one of the promising technology comes density-reachable, density-reachable and density-connected,
in the end in the field of computer science in which it will cluster. Compared with other clustering algorithms, density-
extract the crucial or useful information from massive dataset based clustering technique, such as DBSCAN, has several
or large amount of information. Clustering has play important advantages as follows.
role in the data mining. The process of finding similarities 1. The number of clusters in a data set is not required to be
between data according to the characteristics found in the data input before carrying out the clustering.
and grouping similar data objects into clusters is called 2. The detected clusters can be represented in an arbitrary.
clustering. During the last decades, clustering techniques have 3. Noise or outlier are detected or removed with help of filter.
attracted a lot of attention of researchers for this kind of 4. DBSCAN requires only two parameters which is used find
attention number of clustering algorithms have been proposed the Euclidean and Manhattan distance.
in which DBSCAN is one of them, but the application to large
spatial databases introduces the following requirements: Disadvantage:
1. Minimum number of input parameters: Due to large 1. It is not easy task to calculate the exact initial value of Eps
spatial databases it is not easy to find the initial parameters and MinPts.
like number of clusters, shape and density in advance. 2. Sometime it is very difficult to input the parameters setting.
2. Detect of clusters with arbitrary shape: Because the This paper discuss about improved DBSCAN algorithm from
shape of clusters may be in any random shape. field of data mining. The main conceptual idea of this paper is
3. Good performance should be achieved in very large to compare the number of cluster formed, un-clustered
databases. instances, incorrectly clustered instances, distance change with
The important classes of Clustering are partitioning, help of input parameter and time taken as well as noise ratio of
hierarchical and density-based. To challenge the clustering different dataset. These all dataset are run on WEKA tool.

978-1-4799-1626-9/13/$31.00 2013
c IEEE 42
 The rest of this paper is organized as follows. Section II reduced and the quality of clusters is also improved. In Fast
discusses Literature Survey on clustering techniques. The DBSCAN, objects are sorted by certain dimensional
proposed algorithms for improved DBSCAN algorithm are coordinates. In improved DBSCAN algorithm, global Eps
presented in Section III and proposed Table and Graph are parameter is used. Few or single cluster consisting all object is
reported in section A & B respectively. Section IV describes formed when the range of Eps is small and if the range of Eps
the conclusion and future work and Last Section represented is high many small cluster are generated. The important
the references. advantage of this process is they will reduced the time
complexity and limitation of this method is different density
II. LITERATURE SURVEY clusters is not analyzed.
This section consist the literature review on DBSCAN Tran et al. (2013) [6] has proposed a revised
machine learning algorithm. The main objective is to find the DBSCAN algorithm which become unstable when detecting
benefits and limitations of the DBSCAN algorithm. border objects of adjacent clusters. The final clustering result
El-Sonbaty et al. (2004) [1] has proposed an obtained from DBSCAN depend on the order in which object
algorithm which uses dataset partitioning as a pre-processing are processed in the course of the algorithm run. It retains the
stage. It reduces the number of dataset scan and buffer size key properties of the original DBSCAN algorithm, but in
space is required to keep the partition rather than the whole addition has the potential to improve the clustering results by
dataset. The proposed algorithm can be used for clustering solving the issue of border object. This is achieved by
large datasets and better performance can be obtained. The modifying the expansion step in which core-density-reachable
important advantage is that it is more scalable and can be chains, which contain only core objects, are used for
parallelized easily. The limitation is that results are not clustering.
evaluated on real datasets. C.Havens et al. (2012) [7] has compared the
Duan et al. (2006) [2] has present local DBSCAN efficiency of three different technique aimed to extend fuzzy
algorithm in which appropriate parameters LOFUB, pct and c-means (FCM) clustering. Specially, we compare methods
MinPts and one more point p of the respective cluster is that are based on sampling, incremental technique and
selected. Then all points that are local density reachable from kernelized version of FCM that provide approximations based
the given core point using correct parameters are retrieved. on sampling, including three proposed algorithms. We use
The advantage is that LOF helps in indentify the outliers and loadable and syntactic dataset to conduct the numerical
easily select the appropriate parameters by users. But the experiment that facilitate comparisons based time and space
Cluster analysis is hard in this process. Then all points that are complexity, speed, quality of approximations to batch FCM,
local density reachable from the given core point using correct and assessment of matches between partitions and ground
parameters are retrieved. truth.
Zhang et al. (2007) [3] has shown a Linear DBSCAN S.Vijayalaksmi and M Punithavali (2012) [8] has the
algorithm based on LSH (Locality-Sensitive Hashing) for the paper define the modification of the traditional DBSCAN
purpose of devising main memory algorithm for nearest search algorithm in two manners. The first method uses k-
[4]. The advantage of using LSH is that it reduces the time dimensional tree instead of traditional R-tree algorithm while
complexity and the scale of data. The proposed algorithm second method includes locally sensitive hash procedure to
considered two parts. In the first part, LSH index is built and speed up the process of clustering and increase the efficiency
in the second part clustering is done by the DBSCAN of clustering. The advantage of both approaches are that they
algorithm on the basis LSH retrieval index. The original are unsupervised methods and fully automatic and require no
DBSCAN algorithm cannot handle large scale data but the input from the user.
proposed algorithm is better in handling large scale databases. Glory H Shah (2012) [15] has detected the problem of
The disadvantage of this method is hard to choose the value of clustering, in which cluster are of different size, density and
input parameters. shape. For this DBSACAN clustering algorithm is proposed to
Patwary et al. (2012) [4] has demonstrate a new detect cluster that exists within in a cluster. They evaluated
scalable parallel DBSCAN algorithm using graph algorithmic the result by describing parameters such as number of clusters,
concepts. To construct clusters, a tree based bottom-up unclustered instances as well as incorrectly clustered instances.
approach is used. The disjoint-set data structure is used to For experimental work, they used five different datasets to
break the data access order and to perform the merging evaluate the result.
efficiently. In disjoint-set data structure, two main operations III. PROPOSED ALGORITHM
are used: FIND and UNION. This merging is performed using
master-slave approach where master performs merging In this section, the algorithm DBSCAN (Density Based Spatial
sequentially. The important advantage is that use of master- Clustering of Applications with noise) is designed to discover
salve method which is helps to speed up the process. The main the spatial data clusters with noise. DBSCAN is very sensitive
limitation of this process is it will increase the I/O load and to clustering parameters MinPts and Eps. The steps involved
effects of cost is exits on it. in this algorithm are as follows
Bing Liu (2006) [5] has presented a fast density
based clustering algorithm with which time complexity is

2013 IEEE International Conference in MOOC, Innovation and Technology in Education (MITE) 43
 Table 1. Description of Data sets

Data Set No. of No. of Size Application

Instances Attributes

Lymph 148 19 22 Medical

KB
Sonar 208 61 92 Aerospace
KB
Credit 690 16 34 Banking
Approval KB
Figure.1 Flowchart of Proposed Algorithm Soybean 683 36 198 Agriculture
KB
a) Read Input Dataset Glass 214 10 17.4 Criminal
We have used different type of dataset. There is total KB Investigation
number of 61777 instances with 174 attribute in which 162
attributes are used to evaluate the result. All these instances Pendigits 10992 17 725 Image
are splitting into different dataset. These dataset has been KB processing
downloaded from UCI Machine Learning Repository Site. Adult 48842 15 5089 Census
b) Select WEKA Tool KB Bureau
After read the dataset the next step of our proposed
algorithm to run these all dataset through the WEKA tool of
data mining. Table 2. Detailed Description of Parameters
c) Apply DBSCAN Algorithm
The important step of this algorithm is that, when all Name of the Description
the dataset are loaded in WEKA tool, and then apply the Parameter
DBSCAN algorithm on different dataset with configures the Epsilon Value Maximum radius of the
parameter value. neighborhood.
d) Calculate Parameters Minimum Point Minimum no. of point in an
Before apply DBSCAN algorithm on the datasets Eps-neighborhood of that
user should configure the parameters value which will provide point.
the different result as compare to default value of parameter. Un-clustered The instances which are not
This is very crucial step of our algorithm. Instances form a cluster.
e) Result Fetching Incorrectly Clustered The instances which are not
In this step performance is evaluated and compares Instances formed correct cluster.
the result of dataset with each other in the form of number of
cluster formed, un-clustered instances, incorrectly clustered Time Measure The time in which no. of
instances, time measure and analysis of noise with help of Eps cluster are formed during
and MinPt. clustering.
f) Plot Graphs
No. of Core Point The points that are at interior
After fetching the result, performance is measure
of a cluster.
with graphical representation of different data set.
No. of Border Point The points that are on the
border of cluster
A. EXPERIMENTAL STUDIES
To test the effectiveness of improved DBSCAN
B. PERFORMANCE EVALUATION
algorithm we use different type of datasets. All datasets load
In this experiment, different datasets are normalized
in DBSCAN algorithm and run on WEKA tool of Data
Mining. Also the data sets used in ARFF (attribute Relation according to the work and used with DBSCAN algorithm.
File Format) form. The Detailed Description of all data set and After applying DBSCAN algorithm on different dataset, the
Detailed Description of all parameters are given as shown in performance of each dataset is evaluated and the following
are:
tables.

44 2013 IEEE International Conference in MOOC, Innovation and Technology in Education (MITE)
 As shown in the graph, fig. 3 and fig. 4 it is that using the
selection of input parameter Eps and Minpts, then number of
cluster formed in figure 4 is more as compare to the figure 3
and accuracy in the term of incorrectly clustered instances in
figure 4 is also more. But un-clustered instances in figure 3 are
more as compare to figure 4.

Figure.2 Comparison between Proposed and Existing algorithm

In fig. 2 show the comparison between existing and proposed

work on the basis of number of cluster formed during
experiment on different datasets. It is found that proposed
algorithm contains more cluster as compare to the existing
work. By doing this accuracy of our proposed algorithm is
more than the exiting algorithms.

Figure.5 Computational time taken

In fig. 5 represent the time taken by various type of dataset

which is clearly indicates that soybean has taken more time
than others.

Figure.3 Performance of algorithm on the basis of parameter selection

Figure.6 Result of Adult Dataset

In fig. 6 shows that Adult dataset which have maximum

number of instances in this number of cluster formed is more
as compare to other dataset but it will take more time than
other dataset.

Figure.4 Performance of algorithm on the basis of exact selection of parameter

2013 IEEE International Conference in MOOC, Innovation and Technology in Education (MITE) 45
 REFERENCES
[1] Yasser El-Sonbaty, M. A. Ismail, Mohamed Farouk “An Efficient
Density Based Clustering Algorithm for Large Databases”, IEEE, ICTAI
2004.
[2] Lian Duan, Deyi Xiong, Jun Lee and Feng Guo, “A Local Density Based
Spatial Clustering Algorithm with Noise”, In: Proc. of IEEE
International Conference on Systems, Man, and Cybernetics, Taipei,
Taiwan, October 2006.
[3] Wu, Y. Jou, J. Zhang, X., “A Linear Dbscan Algorithm Based On Lsh”,
Proceedings of the Sixth International Conference on Machine Learning
and Cybernetics, Hong Kong, 19-22 August 2007.
[4] Md. Mostofa Ali Patwary, Diana Palsetia, Ankit Agrawal “A New
Scalable Parallel DBSCAN Algorithm Using Disjoint-Set Data
Structure”, In: Proc. of IEEE International Conference, Salt Lake City,
utah, USA November 2012.
[5] Bing Liu, “A Fast Density Based Clustering Algorithm For Large
Databases”, In: Proc. of IEEE Fifth International Conference on
Machine Learning and Cybernetics, Dalian, August 2006
Figure.7 Computational Time Taken Pendigit Vs Adult [6] Thanh N. Tran, Klaudia Drab, Michal Daszykowski, “Revised
DBSCAN algorithm to cluster data with dense adjacent clusters”,
Chemometrics and Intelligent Laboratory Systems, pp. 92-96, Elsevier,
In fig. 7 indicates that the size of dataset is directly 2013.
proportional to time i.e. larger the size of dataset that will take [7] Timothy C. Havens, Senior Member, IEEE, James C. Bezdek,“Fuzzy c-
more time respectively. Means Algorithms for Very Large Data”, IEEE Transactions on Fuzzy
systems, Vol. 20, No. 6, December 2012.
IV. CONCLUSION AND FUTURE WORK [8] S.Vijayalaksmi, M Punithavali, “A Fast Approach to Clustering Datasets
This paper concludes that before forming the number of using DBSCAN and Pruning Algorithms”, IEEE IJCA 2012.
clusters total number of 61777 instances with 162 attributes is [9] Glory H.Shah, “An Improved DBSCAN, “A Density Based Clustering
Algorithm with Parameter Selection for High Dimensional Data Sets”
divided into different type of dataset. The dataset are real type IEEE 2013.
and these dataset are downloaded from UCI Site. From [10] Chandra. E, Anuradha. V. P, A Survey on Clustering Algorithms for
experimental results and algorithm analysis, the following Data in Spatial Database Management System, International ,n Journal
points can be concluded: The Proposed algorithm can of Computer Applications, Col. 24, June 2011.
effectively analyze the cluster for large dataset. The Improved [11] M. Parimala, D. Lopez, N. C. Senthilkumar, A Survey on Density Based
algorithm is scalable than the density-based algorithm as it Clustering Algorithms for Mining Large Spatial Databases, International
Journal of Advanced Science and Technology, Vol. 31, June 2011.
works on splitting dataset instead of working on whole
[12] M. Rehman and S. A. Mehdi, Comparision of Density-Based Clustering
dataset. The dataset having the total sum of both instances Algorithms, 2005.
and attribute are more, the number of formed clusters as well [13] A. Moreira, M. Y. Santos and S. Carneiro, Density-based clustering
as incorrectly clustered instances is also more. In near future algorithms-DBSCAN and SNN, July 2005.
new modified DBSCAN will be proposed which will used [14] Zeng Donghai. The Study of Clustering Algorithm Based on Grid-
parallel programming to speed up the algorithm and also find Density and Spatial Partition Tree. XiaMen University, PRC, 2006.
the exact initial value of Eps and MinPts parameter for large [15] Chandra. E, Anuradha. V. P, A Survey on Clustering Algorithms for
Data in Spatial Database Management System, International Journal of
dataset. Computer Applications, Col. 24, June 2011.
ACKNOWLEDGMENT [16] M. Parimala, D. Lopez, N. C. Senthilkumar, A Survey on Density Based
Clustering Algorithms for Mining Large Spatial Databases, International
Authors express their heartfelt thanks and gratitude Journal of Advanced Science and Technology, Vol. 31, June 2011.
to their current institutions and those who helped directly or [17] M. Rehman and S. A. Mehdi, Comparision of Density-Based Clustering
indirectly to prepare this manuscript. Authors also express Algorithms, 2005.
sincere thanks to their family members who persistently [18] A. Moreira, M. Y. Santos and S. Carneiro, Density-based clustering
algorithms-DBSCAN and SNN, July 2005.
extended their support and help in whatever way was it
[19] Zeng Donghai. The Study of Clustering Algorithm Based on Grid-
required. Density and Spatial Partition Tree. XiaMen University, PRC, 2006.

46 2013 IEEE International Conference in MOOC, Innovation and Technology in Education (MITE)