You are on page 1of 6

Comparison Between Fuzzy C-means Clustering and

Fuzzy Clustering Subtractive in Urban Air Pollution


Collazo-Cuevas J.I1., Aceves-Fernandez M. A.2, Gorrostieta-Hurtado E2., Pedraza-Ortega
J.C2., Sotomayor-Olmedo A.2, Delgado-Rosas M.2
Universidad del Valle de México1 (UVM) Campus Querétaro Boulevard Villas del Mesón No. 1000 Col.
Juriquilla, Querétaro, México.
Centro de Investigación y Desarrollo en Informática y Telecomunicaciones (CIDIT) Facultad de
Informática de la Universidad Autónoma de Querétaro (UAQ), Av. de las Ciencias s/n, Querétaro México2

collazo013@hotmail.com, marco.aceves@uaq.mx, efren.gorrostieta@uaq.mx, carlos.pedraza@uaq.mx


artemio.sotomayor@uaq.mx, manuel.delgado@uaq.mx

Abstract extending the zero-one membership to the


Clustering is generally associated with membership in the interval (0,1), more flexibility
classification problem. In this contribution, the is introduced [1].
implementation of cluster estimation method as a The second reason for using fuzzy modeling
basis of a fuzzy model identification algorithm is simplifying the problem formulation since this
has been developed. A comparison between two reduces the computational cost.
different clustering techniques is presented, This is due to the fact that a non-fuzzy
Fuzzy C-means clustering and Fuzzy Clustering (crisp) model generally results in an exhaustive
Subtractive. Also, an application in modeling the search in a huge space (because some key
relationship between temperature, humidity and variables can only take values 0 and 1), whereas
PM10 concentration in urban air pollution in in a fuzzy model all the variables are continuous,
Liverpool at northwest of England is presented. so that derivatives can be computed to find the
right direction for the search [5][12].
Keywords. Fuzzy model identification, fuzzy c- Finally fuzzy modeling could be automatic
means, fuzzy clustering subtractive. or semi-automatic process by using clustering
techniques like fuzzy c-means clustering
1. Introduction (FCM)[10] and fuzzy clustering subtractive
Clustering of numerical data forms the basis (FCS) [8][9] where, each cluster center can be
of many classification, system modeling, and used as the basis of a rule that describes the
pattern recognition algorithms [1]. Pattern system behavior.
recognition is a field concerned with machine
recognition of meaningful regularities in noisy or 2. Fuzzy Systems
complex environments. In other words, pattern Since 1985 when the fuzzy model
recognition is search structures in data. methodology suggested by Takagi-Sugeno
Pattern recognition, by its nature, is an [1][2][12], as well known as the TSK model, has
inexact science. To deal with the inexactitude, it been widely applied on theoretical analysis,
is helpful to introduce some "fuzziness" into the control applications and fuzzy modeling.
formulation of the problem. Fuzzy system needs the antecedent and
For example, the boundary between clusters consequence to express the logical connection
could be fuzzy rather than crisp; that is, a data between the input-data and output-data that is
point could belong to two or more clusters with used as a basis to produce the desired system
different degrees of membership. In this way, the behavior [6].
formulation is closer to the real-world problem Fuzzy systems provide a simple, flexible
and therefore better performance may be way to express knowledge by using fuzzy IF-
expected. This is the first reason for using fuzzy THEN Rules [5] and fuzzy modeling provide a
modeling for pattern recognition, in fact, fuzzy framework for modeling complex non-linear
modeling means more flexible modeling-by relations, compared to traditional mathematical

978-1-4244-5353-5/10/$26.00 ©2010 IEEE 174


modeling, fuzzy modeling possesses some calculates a measure of the likelihood that each
distinctive advantages, such as the mechanism of data point would define the cluster center, based
reasoning in human understandable terms, the on the density of surrounding data points.
capacity of taking linguistic information from Consider m dimensions n data point (x1,x2,
human experts and combining it with numerical …, xn) and each data point is potential cluster
data and the ability of approximating complex center, the density function Di of data point at xi
non-linear functions with simple models[1- is given by:
5][12]. A set of n fuzzy rules performs a fuzzy
system. The IF-THEN are the basis of TSK and (3)
any other kind of fuzzy system and are given by:

Ri = IF Y1 is A1 and Y2 is A2 and … THEN Z1 is


B1 and Z2 is B2 … where ra is a positive number. The data point
with the highest potential is surrounded by more
Where Ri is ith fuzzy rule and i pertenecent to set data points. A radius defines a neighbor area then
of n fuzzy rules. In order to automatically the data points, which exceed ra, have no
generate a set of fuzzy rules, clustering influence on the density of data point.
techniques are applied where, each cluster center After calculating the density function of each
can be used as the basis of a rule that describes data point is possible to select the data point with
the system behavior. In this contribution, both the highest potential and find the first cluster
Fuzzy C-means Clustering and Fuzzy clustering center. Assuming that xc1 is selected and Dc1 is its
subtractive are used. density, the density of each data point can be
amended by:
3. The Fuzzy C-Means Clustering
Fuzzy C-Means clustering (FCM) is an (4)
iterative optimization algorithm that minimizes the 2
cost function given by:
n c
The density function of data point which is
J =∑ ∑μ m
ik || x k − v i ||2 (1) close to the first cluster center is reduced, so
k=1 i=1 these data points cannot become the next cluster
center. rb defines an neighbor area where the
Where n is the number of data points, c is the
density function of data point is reduced. Usually
number of clusters, xk is the kth data point, vi is the
constant rb > ra. In order to avoid the overlapping
ith cluster center μik is the degree of membership of
of cluster centers near to other(s) is given by [9]:
the kth data in the ith cluster, and m is a constant
greater than 1 (tipically m=2)[10]. The degree of
membership μik is defined by:
1.5 · (5)
1
μ ik = Once amending the density function of data
c
|| x − vi || 2 /( m −1) points is possible to find the next cluster center.
∑ ( k
j =1 || x k − v j ||
)
(2)
Then repeat this process until find all cluster
centers.
The subtractive clustering is used to
Starting with a desired number of clusters c and an determine the number of clusters of the data
initial guess for each cluster center vi, i = 1,2,3… c, being proposed, and then generates a fuzzy
FCM will converge to a solution for vi that model.
represents either a local minimum or a saddle point
cost function [10]. The FCM method employs
fuzzy partitioning such that each point can belong
5. Fuzzy model construction
to several clusters with membership values between When cluster estimation method is applied
0 and 1. FCM include predefined parameters such to a collection of input/output data, each cluster
as the weighting exponent m and the number of center exemplifies a characteristic behavior of
clusters c. the system. Hence, each cluster center can be
used as the basis of a rule that describes the
4. Fuzzy Clustering Subtractive system behavior. Consider a set of c cluster
* * * *
The subtractive clustering method assumes centers {x1 , x 2 , x 3 ,...x c } in an M-dimensional
each data point is a potential cluster center and space. Let the first N dimensions correspond to

978-1-4244-5353-5/10/$26.00 ©2010 IEEE 175


the input variables and the last M–N dimensional computes the output as a weighted average of the
corresponds to output variables. Each vector x i
* output of each rule [1] [4] [8].
could be decomposed into two component
* * *
vectors y i and z i where y i contains the first N
6. Experimental results
* *
A comparison in the performance of both
elements of x i and z i contains the last M–N fuzzy systems generated by Fuzzy Clustering
elements Subtractive and Fuzzy C-Means was made taking
Then consider each cluster center x i is
* in count the number or membership functions,
represents a fuzzy rule that describes the system rules, and Least Mean Square Error for Airborne
behavior. Given an input vector y, the degree to particulate monitoring in the city of Liverpool in
which rule i is fulfilled is defined by: the Northwest England for different years [11].
The parameters to be taken into account are:
* 2 Humidity and Temperature, these datasets are
μi = e−α ||yy || i (6) provided by the UK National Air Quality
Archive. A total of 13 sites were used with
Where α is the constant defined by [9]. similar characteristics: they used a TEOM
monitoring method (microanalyzer that uses a
microbalance to weigh the amount of particle
4 concentration). The list of sites is shown on
α= (7) Table 1.
ra2
Blackpool Bolton Bradford
and ra is a positive constant[9]. Bury Leeds Liverpool
Manchester Preston Salford
The output vector z is computed by:
c Sheffield Stockport Wigan
∑μ z *
i i
Wirral
Table 1. List of sites where TEOMs are located.
z= i−1
c (8)
∑μ i
It was designed for each case of study two
different FIS models the first using subtractive
i−1
clustering, and the second using 2006 Clustering
then this computational model is in terms of a fuzzy C-Mean, the characteristics of the Subtractive
inference system employing if-then rules following Clustering models are shown in Table 2, and the
the form: characteristics of the Clustering C-Means are
shown in Table 3.
IF Y1 is A1 and Y2 is A2 and … THEN Z1 is B1 and Z2
is B2 … IF-
Site And Membership
Variable THEN
where Y1 is the jth input variable and Z1 is the jth
Year Functions
Rules
output varaiable. A1 is an exponential membership Liverpool
function and B1 is a singleton for the ith rule that is 5 14 14
*
2003
represented by cluster center x i , Aj and Bj are Liverpool
given by: 2004 5 15 15
Liverpool
−α (q−y *ij )2 5 18 18
A j (q) = e (9) 2005
B j = z*ij (10) Liverpool
5 12 12
2006
Liverpool
where y *ij is the jth element of y *i and z*ij is the 2007
5 14 14
*
jth element of z i . This computational scheme is Table 2. Characteristics of FIS models using
equivalent to an inference method that uses Subtractive Clustering.
multiplication as the AND operator, weights the
output of each rule by the firing strength, and

978-1-4244-5353-5/10/$26.00 ©2010 IEEE 176


IF-
Site And Membership
Variable THEN
Year Functions
Rules
Liverpool
2003
5 4 4
Liverpool
2004 5 5 5
Liverpool
2005 5 3 3
Liverpool Figure 2. Membership Functions in FIS model
2006
5 3 3
using Clustering C-Means in Temperature in
Liverpool Liverpool 2004
2007
5 3 3
Table 2: Characteristics of FIS models using As it had shown in the cases, when it used
Clustering C-Means. the Subtractive Clustering algorithm has a lower
computational cost than the C-Means algorithm
As shown in tables 2 and 3 both because it has a higher computational cost,
algorithms create in all the cases the same because in all the cases has more number of
number of rules as membership functions(MF), Membership Functions and rules.
in the case of two o more rules use the same The Least Mean Square Error (LMSE),
median and standard deviation, the algorithm between the raw data, and both FIS models
create a new MF, increasing the computational trough the different years are shown from
cost. Figures 3 to 7.
In the case of using Subtractive
Clustering this algorithm every TSK
Membership function in the same input variable
need to have the same width in difference with
the C-Means algorithm, in Figure 1 is shown an
example of the Membership Functions of the
Temperature and in the Figure 2 is shown the
same input data using C-Means algorithm.

Figure 3. LMSE between Raw data and FIS


models in Liverpool 2003

Figure 1. Membership Functions in FIS model


using Subtractive Clustering in Temperature in
Liverpool 2004

Figure 4. LMSE between Raw data and FIS


models in Liverpool 2004

978-1-4244-5353-5/10/$26.00 ©2010 IEEE 177


complex dynamic systems and provide an
automatic way to generate robust fuzzy systems.
Future work will be involved with minimizing the
system error and try novel clustering techniques.
Also, real-time prediction may be carried out using
sensor networks and embedded systems.

8. Acknowledgements
The authors would like to thank Alan
Charton from The Air Quality Archive hosted by
AEA Energy & Environment, on behalf of the
Figure 5. LMSE between Raw data and FIS
UK Department for Environment, Food & Rural
models in Liverpool 2005
Affairs and the Devolved Administrations (DE-
FRA). Also, the authors would like to
acknowledge the financial support of the
Mexican government via PROMEP funding.

References
[1] Takagi, T., and M. Sugeno. Fuzzy Identification of
Systems and its Application to Modeling and Control.
IEEE Trans. Systems Man and Cybernetics. 15: 116-
132. 1985

Figure 6. LMSE between Raw data and FIS [2] Sugeno, M., and G. T. Kang. Structure
models in Liverpool 2007 Identification of Fuzzy Model. Fuzzy Sets and
Systems. 28: 15-33.1988

[3] Mamdani, E. Advances in the Linguistic Synthesis


of Fuzzy Controllers. Int. Journal in Man-Machine
Studies. 8: 669-678. 1976

[4] Papadakis, S. E., and J. B. Theocharis. A GA-


based Fuzzy Modeling Approach for Generating TSK
Models. Fuzzy Sets and Systems. 131: 121-152. 2002

[5] T. Cox, E. The Fuzzy Systems Handbook:


A Practitioner’s Guide to Building, Using, and
Maintaining Fuzzy Systems. Academic Press. 2002
Figure 7. LMSE between Raw data and FIS
models in Liverpool 2007 [6] Sin, S. K., and De Figueiredo. Fuzzy System
Designing Through Fuzzy Clustering and Optimal
preDefuzzification. Proc. IEEE International
Conference on Fuzzy Systems. 190-195.1993
7. Conclusions and Future work
[7] Gomez, A. F., M. Delgado, and M. A. Vila. About
the Use of Fuzzy Clustering Techniques for
Fuzzy clustering techniques allow the Fuzzy Model Identification. Fuzzy Set and Systems.
automatic generation of fuzzy models and can be 106: 179-188. 1999
utilized to predict Urban Air Pollution. In this
case study, a particular type of harmful particles [8] Chiu S. Fuzzy model identification based on
(PM10) were monitored and a prediction was cluster estimation., Journal of Intelligent and Fuzzy
made using two different fuzzy methods. To Systems; Vol. 2, No. 3, pp. 267–78. September 1994
validate the results, the datasets from other cities
apart from Liverpool, England, were used, [9] Yager, R. and D. Filev, Generation of Fuzzy Rules
showing te feasibility of using both methods to by Mountain Clustering, Journal of Intelligent &
Fuzzy Systems, vol. 2, No. 3, pp. 209- 219. 1994
predict the behavior of such a complex system.
Fuzzy C-means clustering and fuzzy clustering
subtractive are useful techniques to describe

978-1-4244-5353-5/10/$26.00 ©2010 IEEE 178


[10] Bezdek, J. C., Pattern Recognition with Fuzzy
Objective Function Algorithms, Plenum Press, NY,
1981.

[11] M. A. Aceves-Fernandez, J. I. Collazo-Cuevas, E.


Gorrostieta, C. Pedraza-Ortega, J. M. Ramos, S.
Canchola. Prognosis of Urban Air Pollution by using a
Fuzzy Clustering System In Northwest England
journal in Computing Science, ISSN 1870-4069.

[12] Gorrostieta, E., J.C. Pedraza, R.J. Carlos, ”Fuzzy


Modelling of Systems”, Proceedings of 11 th IEEE
International Conference on Methods and Models in
Automation and Robotics MMAR 2005, 29 August- 1
September 2005, Miedzyzdroje, Poland. ISBN 83-
60140-85-5.

978-1-4244-5353-5/10/$26.00 ©2010 IEEE 179

You might also like