Volume 4, Issue 4, April – 2019 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165

Evaluation of BIRCH Clustering Algorithm for

Big Data
Akshatha S R Dr. Niharika Kumar
Master of Technology Associate Professor
Department of Computer Science and Engineering Department of Computer Science and Engineering
B.N.M Institute of Technology B.N.M Institute of Technology
Bengaluru, Karnataka 560050 Bengaluru, Karnataka 560070

Abstract:- Clustering algorithm are as of late additionally an incremental strategy that does not require the
recapturing consideration with the accessibility of vast entire dataset ahead of time.
datasets. Many clustering algorithm don’t scale well
with the expanding dataset and requires the cluster BIRCH already solved the scalability problem for
count parameter which is difficult to assess. In a large dataset. To improve the quality of the cluster, the
BIRCH Clustering algorithm, computing the threshold number of cluster is to be given as an input. The main aim is
parameter of BIRCH from the data is evaluated and the to achieve the quality and the speed of the cluster without
issue of Scalability with increasing dataset sizes and the giving the cluster count. This can be achieved by deleting
cluster count is solved. the global clustering phase which is done at the end of the
BIRCH clustering algorithm. It has four phases: the first
Keywords:- BIRCH Clustering Algorithm, Tree-BIRCH, phase of the algorithm is loading the data into the memory,
Flat Tree. second is condensing tree, third is global clustering and the
last is cluster refinement. The cluster count is needed at
I. INTRODUCTION global clustering step so it is removed and yet preserve the
cluster quality and speed. The other part of the BIRCH
Cluster is a subset of objects which are same or clustering algorithm is called as tree-BIRCH
similar. Data clustering is an unsupervised learning where
the set of data points are grouped in such a way that the data The tree-BIRCH may go wrong in dividing the parent
points in a group are similar to each other than to those in node, splitting of clusters and cluster merging. This is
other groups. The clustering algorithms are k-means solved by computing optimal threshold parameter. The
clustering algorithm, Density-Based Spatial Clustering of threshold parameter is computed by minimum distance and
Applications with Noise (DBSCAN), etc. Many clustering the maximum cluster radius. This is calculated
algorithms don’t scale well with the dynamic (increasing) automatically which is done in Bach and Jordan [9]. Tree-
dataset and the algorithms requires the number of clusters to BIRCH is quick and it is used for the online clustering
be formed as an input. The prediction of the cluster count is which is the elimination if global clustering step. To solve
not so easy where it needs the information, such as data the supercluster splitting, the flat tree is introduced. The flat
exploration, feature engineering and cluster document. tree is the tree where it has a parent node and all the child
nodes are attached to that parent node. This is achieved by
Balanced Iterative Reducing and Clustering Using keeping the branching factor infinite. The parent node will
Hierarchies (BIRCH) is an unsupervised and data mining be having the information about all the child nodes that is,
clustering algorithm which is introduced by Zhang et al [8]. the parent node maintain the triples total number of data
This is the high speed clustering algorithm available. It is a point, linear summation and the square summation of the
data mining algorithm which is used to form a cluster over child node. When new data point is entered, the triples are
categorically immensely colossal data-sets. Birch cluster updated which is built in a clustering feature tree. Tree-
algorithm works on only numerical dataset. Favorable BIRCH clusters faster than the k-means clustering algorithm
position of BIRCH is its ability of clustering with the for the same size of the dataset [2].
increasing records in a dataset, multi-dimensional metric
information indicates in an endeavor create the good quality Dataset is also called as informational index which
clustering for a given arrangement of assets (memory and gathers the information. Each rows in a dataset is seems to
time imperatives). Much of the time, BIRCH just requires a be one record. There are many rows and columns it consists.
solitary output of the database. It is neighborhood in that It is a collection of the records. The informational index
each clustering choice is made without examining all records esteems for every one of the factors. Each esteem is
information focuses and right now existing groups. It known as a datum. The informational index may involve
misuses the perception that information space isn't normally information for at least one individuals, relating to the
consistently possessed and only one out of every odd quantity of columns. The term informational collection may
information point is similarly imperative. It makes full likewise be utilized all the more freely, to allude to the
utilization of accessible memory to infer the finest information in a gathering of firmly related tables,
conceivable sub-clusters while limiting I/O costs. It is comparing to a specific investigation or occasion. Here, it is
approach to aim at datasets from two-dimensional. The

IJISRT19AP110 www.ijisrt.com 245

Volume 4, Issue 4, April – 2019 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
geospatial data is being used for BIRCH clustering node. If the threshold value is small then the cluster is
algorithm. Geospatial data includes the location, size, shape divided and if the value of the threshold is increased then
of the object on the earth like mountain, lake, etc. the clusters are merged. In multi threshold birch the value of
the threshold is dynamic. It keeps changing as the cluster
II. RELATED WORK operator.

The scalability problem is solved in k-mean clustering The clustering algorithm Density-Based spatial
algorithm by parallelization [2]. Here the work is done on clustering of application with noise (DBSCAN) is proposed
two types of k-means, first one is sequential k-means and [6]. The DBSCAN clustering algorithm visits all the data
another is the parallel k-means. In sequential k-means, the points many times. The structure of the clustering is
MacQueen[3] is used where the clusters are formed by arbitrarily shaped because spatial databases are spherical.
performing the computations many times. At the starting This clustering algorithm has three points. Core point,
stage the centroid is selected randomly from those data border point and noise point. Eps represents the radius of the
points. Once this is done, the selected centroid is calculated cluster and minpts is the minimum number of point in the
multiple times to select a perfect centroid where it includes cluster. Consider if the minpts value is 4 and the point in a
all the data points. Euclidean distance equation is used to cluster is greater than the minpts then it is called as core
calculate the centroid. The second is parallel k-means where point. The data point which are less than minpts and those
it make use of Modha [4] which is a distributed memory point are within the radius Eps those are the neighbors of the
multiprocessors. In this method the data points are divided core point are called border point. The noise points are those
equally based upon the processors and then the calculations which are not a core point as well as the border point. This
are done on the divided subset. Once this is done then the clustering algorithm need density as an input parameter.
centroid is calculated on the divided subset. The parallel k- Density parameter gives the maximum number of points a
means runtime is much faster than the sequential k-means. cluster can have within a radius Eps. This clustering
Graphical processing unit is a memory used here which is a algorithm handles the noise and the value of Eps and minpts
shared multiprocessor. Processors are also called as thread of the clusters should be known. If one point of the cluster is
which results in master-slave relationship where the master known then all points is checked either it density reachable
will select the centroid. Labelling step in a Graphical or not. In this algorithm, the global values are used for the
processing unit seems to be more advantageous of the k- Eps and minpts parameter throughout the cluster formatting.
means for huge amount of dataset and more number of The cluster with least density parameter is considered as a
cluster counts. Parallel k-means through Compute Unified good cluster. These clusters are said to be a good candidates
Device Architecture is also taken care. This is tested only for the global values which has no noise in it. The clusters
for the limited available memory and the issue of scalability are combined only if the density of two clusters are near to
is solved by the parallel k-means. each other. The evaluation is done on real data and synthetic
data of SEQUOIA2 000 benchmark.
The improved version BIRCH in the threshold value is
studied from the improved multi threshold birch clustering In hierarchical clustering [7], the cluster is formed in a
algorithm [5]. BIRCH uses clustering feature (CF) in each tree manner. It has two strategy: The first is Agglomerative
clustering feature tree. It is used for the huge amount of clustering and the second is divisive clustering. The
dataset. The clustering feature consists of three attribute that hierarchical agglomerative clustering is an algorithm which
is N, LS and SS. N represents the total number of data is also called as Single-linkage (SLINK) and complete
points in a cluster, LS represents the summation of the data linkage (CLINK) in rare cases.it is a leave to root node
points in a cluster and SS represents the square summation concept. The agglomerative clustering algorithm is a bottom
of the data points in a cluster. The clustering feature is up approach where the clusters are merged and has time
maintained at each node with a static threshold value which complexity O(n3) and needs memory O(n2). The maximum
becomes a problem. In paper [5], another element threshold distance between the data points of the clusters are known
value T is added to the clustering feature. This is done to as CLINK and the minimum distance between the data point
make use of multiple threshold value in an algorithm. In of the clusters are known as SLINK. The mean distance
multi threshold birch algorithm, the clustering feature has between the data points of the clusters are known as average
four attributes are N, LS, SS, T. The threshold value is linkage clustering. The variance is calculated and wen it is
given at the initial stage. When the new data point is increased, the two sub-clusters are combined. The process
entered, it has to check for the nearest node in a clustering of clustering is ended with the small number of clusters. The
feature tree. Once the near leaf node is found then it has to other strategy is Divisive clustering algorithm which is the
check for the threshold value. If the threshold value is not top down approach where the clusters are splits as moves
violated then the newly entered node is assigned to that down O(2n). It is a root to leaves concept. The merging and
nearest node and the threshold value is updated to the the splitting of the cluster are done in greedy manner where
clustering feature and the updated threshold value is the result is presented in dendrogram. A dendogram is a tree
considered as the new threshold value. If the threshold value diagram which is used to arrange the clusters from the
is violated by the entered node, then increase the threshold output of hierarchical agglomerative clustering and divisive
value by multiplying the threshold value with the modified clustering. The BIRCH clustering algorithm and its
factor of the threshold. Once it is multiplied then the application is studied [1].
threshold value becomes larger enough to enter the new

IJISRT19AP110 www.ijisrt.com 246

Volume 4, Issue 4, April – 2019 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
III. PROPOSED MODEL point are of opposite side then the cluster is split into two.
If two points are from different cluster but near to each
The dataset is termed as a geospatial data of zones. other form a single cluster when the given threshold is
Each sample in the database is said to be as an instance. large. Decreased in threshold parameter will also decrease
Computations takes place at each instance of the dataset. in cluster combining. Increase in threshold parameter will
Fig. 1 shows the architecture of the BIRCH where the reduces the clustering splitting. The inadequate clustering
dataset is the first phase and the BIRCH takes only can't be affected by the decision of the threshold value.
numerical values. Some of the geographic information This happens if a cluster covers with two areas having a
system of geospatial data are Natural Earth-Vector, Global place with two non-leaf node in the CF-tree is a height-
map, Land and Ocean Boundaries, Elevation, Weather and balanced tree. This sort of mistake is supercluster splitting.
Climate, Hydrology, Snow, Natural disaster, Land cover, To avoid this supercluster splitting, the flat tree is
Forest geographic information system and etc. The second introduced where all the non-leaf are the children of the
phase is the clustering phase where the tree-Birch and root node. The dataset with huge amount of data with large
clustering feature are exercised. Tree-BIRCH clusters ratio of a cluster and distance to the cluster radius produces
200,000 data points into 2000 clusters in a couple of less error. Fig. 2. Shows cluster radius and the cluster
seconds which leading to faster convergence. The output of distance.
this step gives the cluster data. Tree-BIRCH is the fastest
clustering algorithm available.
The characteristics defined from all the mechanisms
focuses on a keyword called tree-BIRCH, which is the
proposed technique. BIRCH takes three parameter that is,
branching factor, threshold, and cluster count. When the
data point is entered, the clustering feature tree and
hierarchical tree is built. In a hierarchical tree, the each node
seem to be a cluster. Intermediate nodes are the parent
clusters and the leaf nodes are the actual clusters. Branching
factor says how many child nodes a parent can have and if it
exceeds then the parent node has to split up. The dataset Fig 2:- Cluster radius and Cluster distance
used is spatial dataset which is given as an input to the
clustering phase. In clustering phase, the BIRCH clustering To deduce the formula for the ideal limit given two
algorithm will scan the dataset and process to form a conditions: first assume that the branching factor is picked
clusters. the advantages of BIRCH clustering algorithm is it sufficiently extensive for the BIRCH tree to be level, and
takes only single scan of whole data and saves the time of second, that every one of the clusters have roughly a similar
scanning repeatedly as done in other clustering algorithms. number of components. All the cluster has the same number
Once the data is loaded, then next step is the formation of of elements will be the flat tree.
clustering feature where it has the information about the
sub-clusters. Once all the data points are evaluated, the The objective is to acquire the ideal limit parameter as
cluster formation is done. The tree-BIRCH can be used for a function of the maximum radius Rmax of the cluster and
the online clustering algorithm. The radius of the clusters the minimum cluster distance Dmin, where both are known.
should not be more than the threshold value. The underlying presumption is that those two qualities are
either definitely known or simple to get.

IV. IMPLEMENTATION

The radius of the cluster and the distance of the cluster

is calculated based on the clustering feature which is
structured at each node of CF tree.

CF = (N, LS, SS).

Where, N represents the number of data points in a

cluster, LS represents the linear summation of the data
points and SS represents the square summation of the data
points. Based upon this clustering feature the centroid of the
cluster, radius of the cluster and distance between the
Fig 1:- Architecture clusters are calculated.

The Clustering Feature tree formation is done depend N

on the entry of the data points. If the data point of a cluster LS   Y j . (1)
entered in increasing the distance from the center, tree- j 1
BIRCH will probably return only one group. If two data

IJISRT19AP110 www.ijisrt.com 247

Volume 4, Issue 4, April – 2019 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
Where Yj indicates the data points from the dataset. If Dmin is clearly larger than Rmax, it would be
The calculation of LS an SS is shown in the equation 1 and beneficial to increase the threshold beyond. While the lower
equation 2. bound on the threshold is nearly the same for all cluster
distances, the upper bound increases linearly, roughly with
N half the increase of the distance. The following expression is
SS   (Y j ) 2 . (2) used for choosing the threshold.
J 1

T  Dmin  Rmax Where, Dmin ≥Rmax.

Flat tree for the cluster feature tree is achieved by

making the branching factor value infinite or maximum
number which should be more than the number of entries.

V. RESULTS

Fig 3:- CF tree

The structure of the clustering feature for flat tree is

shown ion the Figure 5.3. It shows the flat tree where all the
child nodes (sub-clusters) are assigned to the single parent
node. The following equations are used for the calculation
cluster centroid C, radius R and distance between the
clusters D.

LS
Centroid: C  . (3)
N Fig 4:- Dataset

The centroid of the cluster is calculated by dividing The snapshot of the dataset is shown in the Fig. 3.
the summation of the linear sum of data point with the total This dataset is preprocessed and applied for an algorithm.
number of data points in a cluster. Equation 3 is used for the As the threshold increases, the cluster splitting will be
calculation of cluster centroid. . Equation 4 is to calculate decreased. When the threshold decreases the cluster
the radius of the cluster. combining is avoided. The variation of the threshold of tree
birch and the levelled tree is shown in the Fig. 4 and Fig. 5.


N
j 1
(Y j  C ) 2 When the supercluster splitting is avoided in a flat tree
Radius: R   then the cluster count should be less than the tree-BIRCH as
N
shown in Fig. 6. The flat tree runtime is less than the tree-
N * C  SS  2 * C * LS
2
BIRCH as shown in the Fig. 7.
R . (4)
N

To calculate the distance between the clusters,

consider the clustering feature of two clusters as shown in
equation 5.

CFn = [Nn,LSn,SSn] and CFm = [Nm,LSm,SSm]:

 
Nn Nm
j 1 k 1
(Y j  Z k ) 2
D 
Nn * Nm
N n * SS m  N m * SS n  SS  2 * LS n * LS m
. (5) Fig 5:- Threshold for tree-birch
Nn * Nm

IJISRT19AP110 www.ijisrt.com 248

Volume 4, Issue 4, April – 2019 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
REFERENCES

[1]. T. Zhang, R. Ramakrishnan, M. Livny, BIRCH: a new

data clustering algorithm and its applications, Data
Min. Knowl. Discov. 1(2) (1997) 141–182.
[2]. M. Zechner, M. Granitzer, Accelerating k-means on
the graphics processor via cuda, in: First International
Conference on Intensive Applications and Services,
2009, pp.7–15.
[3]. J.B. Macqueen, Some methods for classification and
analysis of multivariate ob-servations, in: Proceedings
of the Fifth Berkeley Symposium on Math, Statistics,
Fig 6:- Threshold for flat tree and Probability, vol. 1, University of California Press,
1967, pp.281–297.
[4]. Inderjit S. Dhillon and Dharmendra S. Modha. A
dataclustering algorithm on distributed memory
multiprocessors. In Large-Scale Parallel Data Mining,
Lecture Notes in Artificial Intelligence, pp 245–260,
2000.
[5]. N. Ismael, M. Alzaalan, W. Ashour, Improved multi
threshold birch clustering algorithm, Int. J. Artif. Intell.
Appl. Smart Devices 2(1) (2014) 1–10.
[6]. M. Ester, H.-P. Kriegel, J. Sander, X. Xu, A density-
based algorithm for discover-ing clusters in large
spatial databases with noise, in: E. Simoudis, J. Han,
U.M. Fayyad (Eds.), Second International Conference
Fig 7:- Cluster count for the deep tree and leveled tree on Knowledge Discovery and Data Mining, AAAI
Press, 1996, pp.226–231.
[7]. Hastie, Trevor; Tibshirani, Robert; Friedman, Jerome
(2009). "14.3.12 Hierarchical clustering". The
Elements of Statistical Learning (2nd ed.). New York:
Springer. pp. 520–528. ISBN 0-387-84857-6.
Archived from the original (PDF) on 2009-11-10.
Retrieved 2009-10-20.
[8]. Zhang, T.; Ramakrishnan, R.; Livny, M. (1996).
"BIRCH: an efficient data clustering method for very
large databases". Proceedings of the 1996 ACM
SIGMOD international conference on Management of
data - SIGMOD '96. pp. 103–
114. doi:10.1145/233269.233324.
Fig 8:- Runtime [9]. F.R. Bach, M.I. Jordan, Learning spectral clustering,
in: Advances in Neural In-formation Processing
VI. CONCLUSION AND FUTURE WORK Systems, 2004, pp.305–312.

Choosing the cluster count for any clustering

algorithm is uneasy and the information about the dataset is
required. The aim of this project is to cluster the data points
without looking at the dataset. This is done by eliminating
the global clustering phase from the BIRCH algorithm
which is the third phase of the algorithm where the cluster
count is required. To do this the levelled tree mechanism is
introduced where in a tree there will be a single level. It is
also called as flat tree. Here in a levelled tree, all the
children in a tree will have a single parent node. This avoids
the supercluster splitting and the runtime of the proposed
tree is less than the deep tree.

The algorithm is worked for the two dimensional

dataset. In future work, it is suggested to work on multi-
dimension and extended dataset.

IJISRT19AP110 www.ijisrt.com 249