New proximity estimate for incremental update of non uniformly distributed clusters

International Journal of Data Mining & Knowledge Management Process (IJDKP) Vol.3, No.5, September 2013
DOI : 10.5121/ijdkp.2013.3509 91
NEW PROXIMITY ESTIMATE FOR
INCREMENTAL UPDATE OF NON-UNIFORMLY
DISTRIBUTED CLUSTERS
A.M.Sowjanya and M.Shashi
Department of Computer Science and Systems Engineering,
College of Engineering, Andhra University, Visakhapatnam, India
ABSTRACT
The conventional clustering algorithms mine static databases and generate a set of patterns in the form of
clusters. Many real life databases keep growing incrementally. For such dynamic databases, the patterns
extracted from the original database become obsolete. Thus the conventional clustering algorithms are not
suitable for incremental databases due to lack of capability to modify the clustering results in accordance
with recent updates. In this paper, the author proposes a new incremental clustering algorithm called
CFICA(Cluster Feature-Based Incremental Clustering Approach for numerical data) to handle numerical
data and suggests a new proximity metric called Inverse Proximity Estimate (IPE) which considers the
proximity of a data point to a cluster representative as well as its proximity to a farthest point in its vicinity.
CFICA makes use of the proposed proximity metric to determine the membership of a data point into a
cluster.
KEYWORDS
Data mining, Clustering, Incremental Clustering, K-means, CFICA, Non-uniformly distributed clusters,
Inverse proximity Estimate, Cluster Feature.
1. INTRODUCTION
Clustering discovers patterns from a wide variety of domain data, thus many clustering
algorithms were developed by researchers. The main problem with the conventional clustering
algorithms is that, they mine static databases and generate a set of patterns in the form of clusters.
Numerous applications maintain their data in large databases or data warehouses and many real
life databases keep growing incrementally. New data may be added periodically either on a daily
or weekly basis. For such dynamic databases, the patterns extracted from the original database
become obsolete. Conventional clustering algorithms handle this problem by repeating the
process of clustering on the entire database whenever a significant set of data items are added. Let
DS be the original data base (static database) and DS∆ be the incremental database.
Conventional clustering algorithms process the expanded database (SD + ∆ SD) to form new
cluster solution from scratch. The process of re-running the clustering algorithm on the entire

92
dataset is inefficient and time-consuming. Thus most of the conventional clustering algorithms
are not suitable for incremental databases due to lack of capability to modify the clustering results
in accordance with recent updates.
In this paper, the author proposes a new incremental clustering algorithm called CFICA (Cluster
Feature-Based Incremental Clustering Approach for numerical data) to handle numerical data. It
is an incremental approach to partitional clustering. CFICA uses the concept of Cluster Feature
(CF) for abstracting out the details of data points maintained in the hard disk. At the same time
Cluster Feature provides all essential information required for incremental update of a cluster.
Most of the conventional clustering algorithms make use of Euclidean distance ( ED ) between the
cluster representatives ( mean / mode / medoid ) and the data point to estimate the acceptability of
the data point into the cluster.
In the context of incremental clustering while adopting the existing patterns or clusters to the
enhanced data upon the arrival of a significant chunk of data points, it is often required to
elongate the existing cluster boundaries in order to accept new data points if there is no loss of
cluster cohesion. The author has observed that the Euclidean distance (ED) between the single
point cluster representative and the data point will not suffice for deciding the membership of the
data point into the cluster except for uniformly distributed clusters. Instead, the set of farthest
points of a cluster can represent the data spread within a cluster and hence has to be considered
for formation of natural clusters. The authors suggest a new proximity metric called Inverse
Proximity Estimate (IPE) which considers the proximity of a data point to a cluster representative
as well as its proximity to a farthest point in its vicinity. CFICA makes use of the proposed
proximity metric to determine the membership of a data point into a cluster.
2. RELATED WORK
Incremental clustering has attracted the attention of the research community with Hartigan’s
Leader clustering algorithm [1] which uses a threshold to determine if an instance can be placed
in an existing cluster or it should form a new cluster by itself. COBWEB [2] is an unsupervised
conceptual clustering algorithm that produces a hierarchy of classes. Its incremental nature allows
clustering of new data to be made without having to repeat the clustering already made. It has
been successfully used in engineering applications [3]. CLASSIT [4] is an alternative version of
COBWEB. It handles continuous or real valued data and organizes them into a hierarchy of
concepts. It assumes that the attribute values of the data records belonging to a cluster are
normally distributed. As a result, its application is limited. Another such algorithm was developed
by Fazil Can to cluster documents [5]. Charikar et al. defined the incremental clustering problem
and proposed a incremental clustering model which preserves all the desirable properties of HAC
(hierarchical agglomerative clustering) while providing a extension to the dynamic case. [6].
BIRCH (Balanced Iterative Reducing and Clustering using Hierarchies) is especially suitable for
large number of data items [7]. Incremental DBSCAN was presented by Ester et al., which is
suitable for mining in a data warehousing environment where the databases have frequent updates
[8]. The GRIN algorithm, [9] is an incremental hierarchical clustering algorithm for numerical
data sets based on gravity theory in physics. Serban and Campan have presented an incremental
algorithm known as Core Based Incremental Clustering (CBIC), based on the k-means clustering
method which is capable of re-partitioning the object set when the attribute set changes [10]. The
new demand points that arrive one at a time have been assigned either to an existing cluster or a

93
newly created one by the algorithm in the incremental versions of Facility Location and k-median
to maintain a good solution [11].
3. FUNCTIONALITY OF CFICA
An incremental clustering algorithm has to perform the primary tasks namely, initial cluster
formation and their summaries, acceptance of new data items into either existing clusters or new
clusters followed by merging of clusters to maintain compaction and cohesion. CFICA also takes
care of concept-drift and appropriately refreshes the cluster solution upon significant deviation
from the original concept.
It may be observed that once the initial cluster formation is done and summaries are represented
as Cluster Features, all the basic tasks of the incremental clustering algorithm CFICA can be
performed without requiring to read the actual data points ( probably maintained in hard disk )
constituting the clusters. The data points need to be refreshed only when the cluster solution has
to be refreshed due to concept-drift.
3.1 Initial clustering of the static database
The proposed algorithm CFICA is capable of clustering incremental databases starting from
scratch. However, during the initial stages refreshing the cluster solution happens very often as
the size of the initial clusters is very small. Hence for efficiency reasons the author suggests to
apply a partitional clustering algorithm to form clusters on the initial collection of data points
( DS ). The author used the k-means clustering algorithm for initial clustering to obtain k number
of clusters as it is the simplest and most commonly used partitional clustering algorithm. Also k-
means is relatively scalable and efficient in processing large datasets because the computational
complexity is O(nkt) where n is the total number of objects, k is the number of clusters and t
represents the number of iterations. Normally, k<<n and t<<n and hence O(n) is taken as its time
complexity [12].
3.2 Computation of Cluster Feature (CF)
CFICA uses cluster features for accommodating the essential information required for
incremental maintenance of clusters. The basic concept of cluster feature has been adopted from
BIRCH as it supports incremental and dynamic clustering of incoming objects. As CFICA
handles partitional clusters as against hierarchical clusters handled by BIRCH, the original
structure of cluster feature went through appropriate modifications to make it suitable for
partitional clustering.
The Cluster Feature (CF ) is computed for every cluster ic obtained from the k-means algorithm.
In CFICA, the Cluster Feature is denoted as,
CFi = {ni , mi , ݉௜
′
, ܳ௜ ,‫ݏݏ‬௜}
where ni → number of data points,

94
mi → mean vector of the cluster Ci with respect to which farthest points are calculated,
݉௜
′
→ new mean vector of the cluster Ci that changes due to incremental updates,
Qi → list of p-farthest points of cluster Ci
‫ݏݏ‬௜ → squared sum vector that changes during incremental updates.
A Cluster Feature is aimed to provide all essential information of a cluster in the most concise
manner. The first two components ni and mi are essential to represent the cluster prototype in a
dynamic environment. ni will be incremented whenever new data point is added to a cluster. ݉௜
′
,
the new mean is essential to keep track of dynamically changing nature / concept – drift occurring
in the cluster while it is growing. It is updated upon inclusion of a new data point.
The Qi , set of p-farthest points of cluster Ci from its existing mean mi , are used to handle non-
uniformly distributed and hence irregularly shaped clusters; The set of p - farthest points of the ith
cluster, are calculated as follows: First Euclidean distances are calculated between the data points
within cluster ic and the mean of the corresponding cluster mi. Then, the data points are arranged
in descending order with the help of the measured Euclidean distances. Subsequently, the top p -
farthest points for every cluster are chosen from the sorted list and these points are known as the p
- farthest points of the cluster ic with respect to the mean value, mi . Thus a list of p - farthest
points is maintained in the Cluster Feature for every cluster, Ci. These p - farthest points are
subsequently used for identifying the farthest point, qi . ‫ݏݏ‬௜, is the squared sum which is essential
for estimating the quality of cluster in terms of variance of data points from its mean
In general, the variance of a cluster ( σ2
)containing ‘N’ data points is defined as
ߪଶ
=
ଵ
ே
∑ ( ‫ݔ‬௝ − ‫̅ݔ‬)ଶே
௝ୀଵ (
ଵ
ே
∑ ‫ݔ‬௝
ଶ
) − ‫̅ݔ‬ଶே
௝ୀଵ .
In the present context, the error associated with ith
cluster represented by its Cluster Feature, CFi is
calculated as given
ߪ௜
ଶ
= ቀ
ଵ
௡೔
∗ ܵܵ௜ ቁ − ݉௜
ଶ
3.3 Insertion of a new data point
Data points in a cluster can either be uniformly or non-uniformly distributed. The shape of a
uniformly distributed cluster is nearly globular and its centroid is located in the middle
(geometrical middle). Non-uniformly distributed clusters have their centroid located in the midst
of dense area, especially if there is a clear variation in the density of data points among dense and
sparse areas of the cluster. The shape of such clusters is not spherical and the farthest points of a
non-uniformly distributed cluster are generally located in the sparse areas.

95
Fig.1 Uniformly Distributed cluster Fig.2 Non-Uniformly Distributed cluster
Here, refers to entities in a cluster
is the cluster representative ( mean / mode / medoid )
represent p-farthest points of a cluster
farthest point in the vicinity of the entity
new entity to be incorporated
Fig.3
In the above figure 3, let point 1, be located on the sparser side and point 2 be located on the
denser side of the cluster at nearly equal distances from the centroid. Point C is the farthest point
in the vicinity of point 1 and Point A is the farthest point in the vicinity of point 2. Even if the
point 1 is at a slightly larger distance than point 2 to the centroid it is natural to include point 1
into the cluster Ci compared to point 2, due to point 1’s closeness to existing members ( though
farthest points) of the cluster as well as the discontinuity between point 2 to the existing boundary
of the cluster on its side.
The results produced by standard partitional clustering algorithms like K-means are not in
concurrence with this natural expectation as they rely upon Euclidean distance metric (ED) for
discriminating data points while determining their membership into a cluster. So, the author
proposes a new proximity metric called Inverse Proximity Estimate (IPE) to determine the

96
membership of a data point into a cluster considering the data distribution within the cluster
through a Bias factor in addition to normal Euclidean distance as shown below
‫ܧܲܫ‬ ∆௬
(௜)
= ‫ܦܧ‬ (݉௜ , ∆‫)ݕ‬ + ‫ܤ‬
where, ‫ܧܲܫ‬ ∆௬
(௜)
proposed distance metric
‫ܦܧ‬ (݉௜ , ∆‫)ݕ‬ Euclidean distance between the centroid, ݉పሬሬሬሬԦ of the cluster Ci
and the incoming data point, ∆‫.ݕ‬
B Bias factor
Bias is the increment added to the conventional distance metric in view of formation of more
natural clusters and better detection of outliers. It considers the unevenness / shape of the cluster
reflected through a set of p- farthest points to estimate the proximity of new data points to the
cluster.
3.3.1 Bias
Fig.4
In Figure 4, A, B, C are the set of farthest points and it can be observed that they are located in
the sparser areas of the cluster while its centroid is in the midst of dense area. Point C is the
farthest point in the vicinity of point 1 and Point A is the farthest point in the vicinity of point 2.
But the distance between C and point 1 is much smaller than the distance between A and point 2
which is taken as one of the factors that assess the bias and the bias increases with the distance of
the new point to a farther point. So, bias is proportional to the Euclidean distance, ED between
the farthest point ( ‫ݍ‬௜) and the incoming new data point ( ∆‫ݕ‬ ).
∴ ‫ܤ‬ ߙ ‫ܦܧ‬ (‫ݍ‬௜ , ∆‫)ݕ‬
Another aspect to be considered is, as the distance between the centroid and the particular farthest
point in the vicinity of the new point increases, elongation of the cluster with respect to that
farthest point should be discouraged. This is depicted in the Figure 5 below.

97
Fig.5
In the Figure 5, point 1 is in the vicinity of C and Point 3 is in the vicinity of A while both A and
C are farthest points. In this context point 3 is more acceptable than point 1 into the cluster as C is
farther than A to the centroid. The bias increases with the distance of a particular farthest point to
its centroid. Therefore, bias is proportional to the Euclidean distance, ED between the centroid
( ݉పሬሬሬሬԦ ) and the farthest point in the vicinity of the incoming data point ( ∆‫ݕ‬ ).
∴ ‫ܤ‬ ߙ ‫ܦܧ‬ (݉௜ , ‫ݍ‬௜)
Hence, bias is estimated as a product of ED (‫ݍ‬௜ , ∆‫)ݕ‬ , ‫ܦܧ‬ (݉௜ , ‫ݍ‬௜) mathematically from the
above equations. Therefore bias is expressed as,
‫ܤ‬ = [ ‫ܦܧ‬ (‫ݍ‬௜ , ∆‫)ݕ‬ ∗ ‫ܦܧ‬ ( ݉௜ , ‫ݍ‬௜) ]
3.3.2 Proposed Proximity Metric – Inverse Proximity Estimate (IPE)
The authors have devised a new proximity estimate ‫ܧܲܫ‬ ∆௬
(௜)
to determine the acceptability of an
incoming data point ∆‫ݕ‬ into a possibly non-uniformly distributed cluster represented by its mean
݉௜ which is calculated as follows
‫ܧܲܫ‬ ∆௬
(௜)
= ‫ܦܧ‬ (݉௜ , ∆‫)ݕ‬ + ‫ܤ‬
Substituting B from the above, now (‫ܧܲܫ‬ ∆௬
(௜)
) is estimated as
‫ܧܲܫ‬ ∆௬
(௜)
= ‫ܦܧ‬ (݉௜ , ∆‫)ݕ‬ + [ ‫ܦܧ‬ (‫ݍ‬௜ , ∆‫)ݕ‬ ∗ ED (݉௜ , ‫ݍ‬௜)]
For clusters with uniformly distributed points, the usual distance measures like Euclidean
distance hold good for deciding the membership of a data point into a cluster. But there exist
applications where clusters have non-uniform distribution of data points. In such cases, the new
distance metric proposed above will be useful. The inverse proximity estimate, ‫ܧܲܫ‬ ∆௬
(௜)
better
recognizes the discontinuities in data space while extending the cluster boundaries as explained
below.
3.3.3 Extendibility of the cluster boundary towards Sparse area Vs Dense area
Extendibility of a cluster to include a new point towards sparse area is based on a set of farthest
points defining its boundary around the sparse area.

98
The points that can be included in the region nearer to the dense area by extending the boundary
of the cluster towards dense area are much nearer to the centroid of this cluster only, based on the
original distance metric, ED as well as the inverse proximity estimate, ‫ܧܲܫ‬ ∆௬
(௜)
. In other words
such data points that are outside the present cluster boundary are closer to the centroid compared
to existing farther members.
However, an external point which is located towards dense area at an equal distance from the
centroid as the farthest points is clearly separated from the dense area with a discontinuity in
between. The inverse proximity estimate, ‫ܧܲܫ‬ ∆௬
(௜)
does not allow extension of the boundary of
the cluster on the denser side, if there is a discontinuity between dense area and the new point,
whereas such discontinuity may be overlooked by the original distance metric, ED.
Let point 5 and point 6 are equidistant from the centroid of the cluster. As shown in the Figure 6,
point 5 is towards its sparse area and point 6 is towards its dense area.
Fig. 7
Though point 6 is at an equal distance to centroid as point 5, point 6 is not acceptable into the
cluster because of the discontinuity.
The only cluster members in the vicinity of point 6, with a possible discontinuity between itself
and the cluster boundary, are located in the dense area. Hence, none of them are considered as
farther points of the cluster. Since the sparse area containing farther points is away from point 6,
Euclidean distance between the farthest point and the incoming new data point, ‫ܧ‬஽ (‫ݍ‬௜ , ∆‫)ݕ‬ is
high for point 6 compared to point 5, increasing the value of bias for point 6. This in turn
increases the inverse proximity estimate, ‫ܧܲܫ‬ ∆௬
(௜)
for point 6 compared to point 5, thereby,
reducing the acceptability of point 6 into the cluster.
3.4. Incremental Clustering Approach with CFICA
CFICA uses the inverse proximity estimate, ‫ܧܲܫ‬ ∆௬
(௜)
for effectively identifying the appropriate
cluster of an incoming data point y∆ . In other words, it estimates the proximity of the incoming
point to a cluster based on the cluster centroid (mean ݉௜), farthest point in the vicinity of the
incoming point ( qi) and the incoming data point (∆y).
For each cluster ic , the Euclidean distance ED is calculated for the following pairs of points:
centroid and incoming point (݉௜, ∆‫,)ݕ‬ farthest point in the vicinity of the incoming point and
incoming point(‫ݍ‬௜ , ∆‫,)ݕ‬ centroid and farthest point in the vicinity of the incoming point (݉௜ , ‫ݍ‬௜).

99
Upon the arrival of a new data point y∆ to the existing database DS which is already clustered
into C = {C1, C2, …….., Ck} clusters, its distance, ‫ܧܲܫ‬ ∆௬
(௜)
to ith
cluster for all i = 1 to k is
calculated using the Equation specified in section 2.3.2
‫ܧܲܫ‬ ∆௬
(௜)
= ‫ܦܧ‬ (݉௜ , ∆‫)ݕ‬ + [ ‫ܦܧ‬ (‫ݍ‬௜ , ∆‫)ݕ‬ ∗ ED (݉௜ , ‫ݍ‬௜)]
where, ‫ܦܧ‬ (݉௜ , ∆‫)ݕ‬ Euclidean distance between the points im and y∆
‫ܦܧ‬ (‫ݍ‬௜ , ∆‫)ݕ‬ Euclidean distance between the points iQ and y∆
ED (݉௜ , ‫ݍ‬௜) Euclidean distance between the points im and iQ
3.5 Finding the farthest point in the vicinity of incoming data point y∆
In order to find the farthest point qi ( data point in Qi ) which is in the vicinity of the incoming
data point y∆ , the Euclidean distance is calculated for that data point y∆ to each of the p -
farthest points of that cluster and the data point with minimum distance is designated as qi
thereby, for each cluster, the point having minimum distance is taken as the farthest point qi..
3.6 Finding the appropriate cluster
The proximity metric, IPE is used to find the appropriate cluster for the new data point, y∆ . The
new data point, y∆ is assigned to the closest cluster only if the calculated proximity, ‫ܧܲܫ‬ ∆௬
(௜)
is
less than the predefined threshold value, λ. Otherwise, the data point y∆ is not included in any of
the existing clusters, but it separately forms a new singleton cluster. In such a case, the number of
clusters is incremented by one.
3.7 Updating of Cluster Feature
From the above section, it can be seen that, whenever a new data point, ∆‫ݕ‬ is added to the
existing database, that new data point may be included into any of the existing clusters or it may
form a new cluster. So after the new point gets inserted, updating of CF is important for further
processing.
Case 1: Inclusion of new data point into any of the existing clusters
Whenever a new data point, ∆‫ݕ‬ is included into an already existing cluster Ci its cluster feature
(CFi) is updated without requiring the original data points of Ci and hence supports incremental
update of the clustering solution.
In particular, the ni, ݉ప
′ ,ሬሬሬሬሬሬሬԦ and ‫ݏݏ‬௜ fields of CFi are updated upon the arrival of a new data point ∆‫ݕ‬
into the cluster Ci as given below:

100
(1) ni = ni + 1
(2) ݉௜
′
=
௡೔ ∗ ௠೔
′
ା∆௬
௡೔
(3) ‫ݏݏ‬௜ = ‫ݏݏ‬௜ + squared components of ∆‫.ݕ‬
However, the Qi representing the p-farthest points and the centroid of the previous snapshot ݉௜
′
,
were kept without any changes until the next periodical refresh.
Case 2: New data point forms a singleton cluster
If the new data point forms a new cluster separately, the cluster feature (CFi ) has to be computed
for the new cluster containing the data point y∆ . CFi for the new cluster contains the following
information:
(1) Number of data points. Here n=1, as it is a singleton cluster,
(2) Mean of the new cluster; In this case, mi = ݉௜
′
= ߂‫ݕ‬
(3) The list of p-farthest points includes ߂‫ݕ‬ only.
(4) Squared sum is squared sum of components of ߂‫.ݕ‬
Finding the appropriate cluster to incorporate the new data point y∆ and updating of the cluster
feature after adding it to the existing cluster or forming a separate cluster are iteratively
performed for all the data points in the incremental database DS∆ .
3.8 Merging of closest cluster pair
Once the incremental database DS∆ is processed with CFICA, the need to merge the closest
clusters may arise. A merging strategy is used to maintain reasonable number of clusters with
high quality. Therefore, merging is performed when the number of clusters increases beyond ‘k’
( k in k-means) while ensuring that increase in variance which indicates error is minimum due to
merging. It is intuitive to expect an increase in the error with the decrease in the number of
clusters. A closest cluster pair is considered for merging if only the Euclidean distance between
the centroids of the pair of clusters is smaller than user defined merging threshold (θ).
The procedure used for the merging process is described below:
Step 1: Calculate the Euclidean distance ED between every pair of cluster centroids (݉పሬሬሬሬԦ).
Step 2: For every cluster pair, with Euclidean distance, ED less than the merging threshold
value, ߠ ( ED ≪ ߠ ), find increase in variance (ߪଶ
) as described in section 3.2.
Step 3: Identify the cluster pair (Ci, Cj) with minimum increase in variance.

101
Step 4 : Merge Ci and Cj to form the new cluster, Ck and compute the Cluster Feature for Ck
and delete Ci and Cj along with their CF’s.
Step 5 : Repeat steps 1 to 4 until no cluster pair is mergable or until the value of ‘k’ is adjusted.
Computing Cluster Feature of merged cluster
After merging the closest cluster pair, now the Cluster Feature ( CF ) has to be computed for the
merged cluster. The CF of the merged cluster is calculated by:
(1) Adding the number of data points in both the clusters
(2) Calculating the mean of the merged cluster
(3) Finding the p-farthest points of the merged cluster
(4) Incrementing the squared sum.
In particular, when the ith
cluster with CFi = { ni, mi , mi
’
, Qi , ssi} and jth
cluster with CFj = { nj ,
mj , mj
’
, Qj ,ssj } are merged to form kth
cluster, its cluster feature CFk is calculated as given below:
1) nk = ( ni + nj )
2) ݉௞ሬሬሬሬሬԦ =
(୬౟ ∗ ௠ഢሬሬሬሬሬԦ ) ା (୬ౠ ∗ ௠ണሬሬሬሬሬሬԦ )
௡ೖ
3) ݉௞
′ሬሬሬሬሬԦ = ݉௞ሬሬሬሬሬԦ
4) p-farthest points are selected from farthest points in qi and qj .
5) ssk = ssi + ssj
Hence the Cluster Feature of the new cluster ‘k’ which is formed by merging two existing clusters
‘i’ and ‘j’ is determined as a function of CFi and CFj.
CFk = f ( CFi , CFj)
Thus Cluster Feature provides the essence of the clustered data points thereby avoiding explicit
referencing of individual objects of the clusters which may be maintained in the external memory
space.
3.9 Need for Cluster Refresh
The addition of new data points into some of the existing clusters naturally results in change of
mean. So, the set of p-farthest points of cluster Ci, need not be the p-farthest points of cluster ‫ܥ‬௜
′
,
modified version of Ci. Whether or not, the p-farthest points of the cluster are the same, depend
upon the recently added data points. Ideally, a new set of p-farthest points have to be computed
for the incremented cluster ‫ܥ‬௜
′
. Calculating p-farthest points again, every time the cluster gets
updated, is not an easy task as it involves recalculation of the Euclidean distance, ED for every
data point in the incremented cluster ‫ܥ‬௜
′
with the updated mean, ݉௜
′
.

For pragmatic reasons, it was suggested to refresh the
from its original value indicating concept drift [Chen H.L et al. 2009]. After processing a new
chunk of data points from ∆SD the deviation in mean is calculated
Those clusters with deviation in mean greater than
clusters with significant concept
refresh involves finding a new set of p
clusters, the same set of p-farthest points along with the old mean value, m
be noted that mi in the CFi always represents the centro
identified. Hence needs to be changed whenever new set of p
above process of cluster refresh is applicable only to the incremented cluster C
of additional data points
4. EXPERIMENTAL ANALYSIS
CFICA has been implemented using the Iris [13], wine [14] and yeast [15] datasets from the UCI
machine learning repository. All the datasets were preprocessed. Iris and Wine datasets do not
have any missing values but Yeast dataset has some missing attribute values and such instances
were ignored. After preprocessing, yeast dataset had 1419 instances. Normally datasets used for
the purpose of analysis may contain too many attributes, which may or may not be relevant.
Therefore dimensionality reduction has been done on Wine and Yeast datasets as they contain a
considerable number of attributes. So the number of attributes came down from 13 to 3 for wine,
from 8 to 5 attributes for yeast datasets.
Fig 1. Purity of Wine dataset before and after dimensionality reduction
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
4
of Data Mining & Knowledge Management Process (IJDKP) Vol.3, No.5, September 2013
ragmatic reasons, it was suggested to refresh the CFi, only in case it deviates significantly
the deviation in mean is calculated as follows
Deviation in mean =
௠೔
′
ି ௠೔
௠೔
Those clusters with deviation in mean greater than δ, based on a threshold are identified as
clusters with significant concept – drift and hence needs to be refreshed. The process of c
refresh involves finding a new set of p-farthest points for such clusters. For the remaining
farthest points along with the old mean value, mi is maintained. It may
always represents the centroid based on which the p-farthest points are
identified. Hence needs to be changed whenever new set of p-farthest points are identified. The
above process of cluster refresh is applicable only to the incremented cluster Ci
’
due to inclusion
NALYSIS
Yeast dataset has some missing attribute values and such instances
erefore dimensionality reduction has been done on Wine and Yeast datasets as they contain a
from 8 to 5 attributes for yeast datasets.
Fig 1. Purity of Wine dataset before and after dimensionality reduction
8
12
16
20
Before
After
102
, only in case it deviates significantly
, based on a threshold are identified as
drift and hence needs to be refreshed. The process of cluster
farthest points for such clusters. For the remaining
is maintained. It may
farthest points are
farthest points are identified. The
due to inclusion
Yeast dataset has some missing attribute values and such instances
erefore dimensionality reduction has been done on Wine and Yeast datasets as they contain a
Before
After

Fig 2. Purity of Yeast dataset before and after dimensionality reduction
It can be seen from the above figures that the cluster purity increases after dimensionality
reduction has been done.
4.1 Cluster formation
The above datasets are processed dynamically by dividing the data points in each dataset into
chunks. For example, the Iris dataset is divided into 4 chunks (Iris1 consisting of 75 instances,
Iris2, Iris3 and Iris4 consisting of 25 instances each). Initially, the first chunk of data points is
given as input to the k-means algorithm for initial clustering i.e Iris1. It generates k number of
clusters. Then, the cluster feature is computed for those k initial clusters. The nex
points is input to the new approach incrementally. For each data point from the second chunk, the
Inverse Proximity Estimate (IPE) is computed and the data points are assigned to the
corresponding cluster if the calculated distance measure
value, λ. Otherwise, it forms as a separate cluster. Subsequently, the cluster feature is updated for
each data point. Once the whole chunk of data points is processed, the merging process is done if
only the Euclidean distance between the centroids of the pair of clusters is smaller than user
defined merging threshold, θ. Here,
obtained from the merging process and the cluster solution is updated incrementally
receiving the later chunks of data Iris3 and Iris4. Similarly, wine and yeast datasets are also
divided into 4 chunks as follows: wine1
wine4 - 28 instances, yeast1 - 700 instances, yeast 2
yeast4 - 168 instances and handled as above.
4.2 Metrics in which performance is estimated
Validation of clustering results is important. Therefore, the Purity Measure has been used to
evaluate the clustering results ob
to a single class. The purity measure described in [16] [17] has been used for evaluating the
performance of CFICA. The evaluation metric used in CFICA is given below,
0
0.2
0.4
0.6
0.8
4 8 12
Fig 2. Purity of Yeast dataset before and after dimensionality reduction
ing of 25 instances each). Initially, the first chunk of data points is
means algorithm for initial clustering i.e Iris1. It generates k number of
clusters. Then, the cluster feature is computed for those k initial clusters. The next chunk of data
corresponding cluster if the calculated distance measure is less than the predefined threshold
. Otherwise, it forms as a separate cluster. Subsequently, the cluster feature is updated for
distance between the centroids of the pair of clusters is smaller than user
θ. Here, λ = 10 and θ = 4.Finally, the set of resultant clusters are
obtained from the merging process and the cluster solution is updated incrementally
divided into 4 chunks as follows: wine1 - 100 instances, wine2 and wine3 - 25 instances each,
700 instances, yeast 2 – 350 instances, yeast3 - 200 instances,
168 instances and handled as above.
4.2 Metrics in which performance is estimated
evaluate the clustering results obtained. A cluster is called a pure cluster if all the objects belong
performance of CFICA. The evaluation metric used in CFICA is given below,
Purity =
ଵ
ே
∑ ܺ௜
்
௜ୀଵ
16 20 22 25 28 30
Before
After
103
ing of 25 instances each). Initially, the first chunk of data points is
means algorithm for initial clustering i.e Iris1. It generates k number of
t chunk of data
is less than the predefined threshold
. Otherwise, it forms as a separate cluster. Subsequently, the cluster feature is updated for
distance between the centroids of the pair of clusters is smaller than user
= 4.Finally, the set of resultant clusters are
obtained from the merging process and the cluster solution is updated incrementally upon
25 instances each,
200 instances,
tained. A cluster is called a pure cluster if all the objects belong
Before
After

104
where, N Number of data points in the dataset
T Number of resultant cluster
iX Number of data points of majority class in cluster i
The purity of the resultant clusters is calculated by changing the k-value (order of initial
clustering) and is graphically presented in Figures 3 to 5.
Fig 3. Purity vs. number of clusters ( k ) for Iris dataset.
Fig 4. Purity vs. number of clusters ( k ) for Wine dataset.
Fig 5. Purity vs. number of clusters ( k ) for Yeast dataset.

105
The above results demonstrate that CFICA performs better than BIRCH in terms of purity.
BIRCH algorithm is unable to deliver satisfactory clustering quality if the clusters are not
spherical in shape because it employs the notion of radius or diameter to control the boundary of
a cluster.
5. THE PSEUDO-CODE FOR CFICA
The various steps constituting the proposed algorithm CFICA is listed out in the form of pseudo-
code. CFICA accepts the clustering solution for existing database SD in the form of cluster
features as input and incrementally updates the set of cluster features in ∆SD. in accordance with
newly arrived chunk of data points.
Input
ܵ஽ → Initial set of CF’s for SD.
∆ܵ஽ → Set of data points added to SD
K → Number of clusters
θ → Predefined merging threshold value
δ → Allowed deviation in mean
λ → User defined radius threshold
Output
Set of CF’s { CF1, CF2, . . . . . . , CFK } for ( SD.+ ∆SD)
Variables
∆‫ݕ‬ → New data point
‫ܨܥ‬௜ → Cluster Feature of ith
cluster
‫ܧܲܫ‬ ∆௬
(௜)
→ Inverse Proximity Estimate between ∆‫ݕ‬ and ith
cluster
ED → Euclidean distance
݊௜ → Number of data points
݉௜ → Mean vector of the cluster
݉௜
′
→ New mean vector of the cluster
ܳ௜ → List of p-farthest points of ith
cluster
‫ݍ‬௜ → The data point in ܳ௜ which is closest to ∆‫ݕ‬
ssi → Squared sum vector

106
Algorithm
1) Compute Cluster Feature for every cluster ‫ܥ‬௜
CFi = { ni, mi, mi
'
, Qi, ssi}
a = k;
2) For each data point ∆‫ݕ‬ ߳ ∆ܵ஽
a) For i from 1 to a
Calculate Inverse Proximity Estimate of ∆‫ݕ‬ to the ith
cluster
‫ܧܲܫ‬ ∆௬
(௜)
= ‫ܦܧ‬ (݉௜ , ∆‫)ݕ‬ + [ ‫ܦܧ‬ (‫ݍ‬௜ , ∆‫)ݕ‬ ∗ ED (݉௜ , ‫ݍ‬௜)]
b) Find suitable cluster j = arg mini { ‫ܧܲܫ‬ ∆௬
(௜)
}
if ( ‫ܧܲܫ‬ ∆௬
(௜)
< λ )
{
insert ∆‫ݕ‬ into its closest cluster j and update its CF as follows:
CFj = { (nj +1), mj,
௡ೕ∗ ௠ೕା ∆௬
௡ೕା ଵ
, qj,(SSi +squared components of ∆‫})ݕ‬
if ( deviation in mean =
|௠ೕ
′
ି ௠ೕ|
|௠ೕ|
> ߜ)
Read the data points of jth
cluster to recompute CFj.
}
else create a new cluster
{
increment a
insert ∆‫ݕ‬ into ath
cluster
CFa = {1, ma , ma
'
, ∆y, squared components of ∆y}
}

107
3) If ( a > k ) merge cluster pair
(i) Compute ED between every pair of cluster centroids based on ݉௜.
(ii) For cluster pairs with ( ED≪ ߠ ), find increase in variance.
(iii) Merge the cluster pair with minimum increase in variance
Decrement a
(iv) Recalculate CF of merged cluster
CFk = { ( ni + nj ),
(୬౟ ∗ ୫౟) ା (୬ౠ ∗ ୫ౠ)
௡ೖ
, ݉k = ݉௞
′
, qk,, SSk = SSi + SSj }
Find its ED to other clusters.
(v) Repeat steps (i) to (iv) until ( a = k )
4) Print cluster solution as the set of cluster features.
Wait for the arrival of new chunk of data points upon which call CFICA
again.
6. CONCLUSIONS
An incremental clustering algorithm called Cluster Feature-Based Incremental Clustering
Approach for numerical data (CFICA) which makes use of Inverse Proximity Estimate to handle
entities described in terms of only numerical attributes was developed and evaluated. Cluster
Feature while being compact includes all essential information required for maintenance and
expansion of clusters. Thus CFICA avoids redundant processing which is the essential feature of
an incremental algorithm. The performance of this algorithm, in terms of purity is compared with
the state of art incremental clustering algorithm namely BIRCH on different bench mark datasets.
REFERENCES
[1] Hartigan, J.A. 1975. Clustering Algorithms. John Wiley and Sons, Inc., New York, NY.
[2] Fisher D., “Knowledge acquisition via incremental conceptual clustering,” Machine Learning,
vol. 2, 1987, pp.139-172.
[3] Fisher , D., Xu, L., Carnes, R., Rich, Y., Fenves, S.J., Chen, J., Shiavi, R., Biswas, G., and
Weinberg, J. 1993. Applying AI clustering to engineering tasks. IEEE Expert 8, 51–60.
[4] J. Gennary, P. Langley, and D. Fisher, “Models of Incremental Concept Formation,” Artificial
Intelligence Journal, vol. 40, 1989, pp. 11-61.

108
[5] Fazil Can, “Incremental Clustering for Dynamic Information Processing”, ACM Transactions on
Information Systems, April 1993, Vol. 11, No. 2, pp. 143-164.
[6] M. Charikar, C. Chekuri, T. Feder, and R. Motwani, “Incremental clustering and dynamic
information retrieval,” 29th Symposium on Theory of Computing, 1997, pp. 626—635.
[7] Tian Zhang , Raghu Ramakrishnan , Miron Livny, “BIRCH: an efficient data clustering method
for very large databases”, Proceedings of the ACM SIGMOD International. Conference on
Management of Data, pp 103-114, 1996.
[8] Ester, M., Kriegel, H., Sander, J., Xu X., Wimmer, M., “Incremental Clustering for Mining in a
Data Warehousing Environment”, Proceedings of the 24th International. Conference on Very
Large Databases (VLDB’98), New York, USA, 1998, pp. 323-333.
[9] Chien-Yu Chen, Shien-Ching Hwang, and Yen-Jen Oyang, "An Incremental Hierarchical Data
Clustering Algorithm Based on Gravity Theory", Proceedings of the 6th Pacific- Asia
Conference on Advances in Knowledge Discovery and Data Mining, 2002, pp: 237 – 250.
[10] Serban G., Campan A., "Incremental Clustering Using a Core-Based Approach", Lecture Notes
in Computer Science, Springer Berlin, Vol: 3733, pp: 854-863, 2005.
[11] Fotakis D., "Incremental algorithms for Facility Location and k-Median", Theoretical Computer
Science, Vol: 361, No: 2-3, pp: 275-313, 2006.
[12] Jain A.K., Murthy M.N., Flynn P.J., “Data Clustering: A Review”, ACM Computing Surveys,
September 1999, Vol. 31, No.3, pp. 264 – 323.
[13] Iris dataset : http://archive.ics.uci.edu/ml/datasets/Iris
[14] Wine dataset : http://archive.ics.uci.edu/ml/datasets/Wine
[15] Yeast dataset : http://archive.ics.uci.edu/ml/datasets/Yeast
[16] Huang Z, “Extensions to the k-Means Algorithm for Clustering Large Data Sets With
Categorical Values”, Data Mining and Knowledge Discovery, 1998, Vol. 2, pp. 283–304.
[17] Xiaoke Su, Yang Lan, Renxia Wan, and Yuming Qin, “A Fast Incremental Clustering
Algorithm”, Proceedings of the 2009 International Symposium on Information Processing
(ISIP’09), 2009, pp. 175-178.
[18] Sowjanya A.M. and Shashi M., “Cluster Feature-Based Incremental Clustering Approach
(CFICA) For Numerical Data”, IJCSNS International Journal of Computer Science and Network
Security, Vol.10, No.9, September 2010.

109
AUTHORS
A.M.Sowjanya received her M.Tech. Degree in Computer Science and technology from
Andhra University. She is presently working as an Assistant Professor in the department
of Computer Science and Systems Engineering, College of Engineering (Autonomous),
Andhra University, Visakhapatnam, Andhra Pradesh, India. She submitted her Ph.D in
Andhra University. Her areas of interest include Data Mining, incremental clustering and
Database management systems.
M.Shashi received her B.E. Degree in Electrical and Electronics and M.E.Degree in
Computer Engineering with distinction from Andhra University. She received Ph.D in
1994 from Andhra University and got the best Ph.D thesis award. She is working as a
professor of Computer Science and Systems Engineering at Andhra University, Andhra
Pradesh, India. She received AICTE career award as young teacher in 1996. She is a co-
author of the Indian Edition of text book on “Data Structures and Program Design in C”
from Pearson Education Ltd. She published technical papers in National and
International Journals. Her research interests include Data Mining, Artificial intelligence, Pattern
Recognition and Machine Learning. She is a life member of ISTE, CSI and a fellow member of Institute of
Engineers (India).

New proximity estimate for incremental update of non uniformly distributed clusters

More Related Content

What's hot

Similar to New proximity estimate for incremental update of non uniformly distributed clusters

Recently uploaded

New proximity estimate for incremental update of non uniformly distributed clusters