Sci.Int.(Lahore),27(3),2055-2061,2015 ISSN 1013-5316; CODEN: SINTE 8 2055 PROPOSED DYNAMIC CLUSTERING BASED ON DISCRETE PSO-GA Hamid Tabatabaee 1,* ,Maliheh Hassan Pour2,Ameneh fakhraee3, ,Mohammad Reza Baghaeipour4 1,* Young Researchers and Elite Club, Quchan Branch, Islamic Azad University, Quchan, Iran. 3,2 Department of Computer Engineering ,Mashhad Branch, Islamic Azad University, Mashhad, Iran. 4 Department of Computer Engineering, Ferdows Higher Education Institute, Mashhad, Iran. ABSTRACT: Data clustering is the process of identifying clusters or natural groups based on similarity measures. Despite the improved clustering analysis algorithms, most of the clustering algorithms likewise require number of clusters as an input parameter. Hence, this paper describes a new method of clustering, called dynamic clustering based on discrete particle swarm optimization and genetic algorithms (DPSO-GA). The proposed algorithm automatically determines the optimal number of clusters and at the same time clusters the data set with minimal user intervention. In this method, the discrete PSO algorithm as well as genetic are employed with a new display solutions which are of variable length to determine the number of clusters and initial cluster centers for K-Means algorithm to improve its performance. The results obtained from five UCI evaluation datasets show that the proposed algorithm has higher performance in comparison with other methods. Keywords: Clustering, Particle swarm optimization, Genetic algorithms. 1. INTRODUCTION Clustering is one of the important category of Unsupervised learning methods for grouping objects into groups distinct from the clusters [ 1]. Cluster grouping is usually defined as the objects grouped into one cluster are similar according to the given criteria while they are different with the objects from other clusters according to the criteria given [ 1] . In recent years, clustering is widely used in various fields such as pattern recognition, machine learning and market analysis [ 2 ]. Analysis of clustering is a significant technique in data mining [ 3 ] . The primary objective of clustering is to put similar data within a cluster and then collect the highsimilarity clusters [ 3 ]. Through clustering, valuable information such as the distribution and characteristics of data which come from a very large data is obtained and hidden information is extracted by reducing the complexity of the data. Thus, the results of clustering can be effectively applied to solve problems or to help with decision making. Four types of clustering algorithms have been identified: the first group of clustering algorithms is based on the idea that the neighborhood data should be placed in the same cluster. As it is stated in [ 4 ] this type of algorithms is robust to recognize the clusters of any shape, but it fails when there is a small spatial separation between the clusters. The second class of clustering algorithms has been made by means of the dynamic changes in the cluster to make the ultimate solution. These algorithms are the most popular clustering algorithms contains k-means [ 5 ] and the other category is based on model-based clustering [ 6 ]. The fourth class of clustering algorithms characterized by different methods which optimize the different attributes of data set [ 7 , 8 ]. So far, the speed and efficiency of data processing has been the focus of clustering development and the most clustering methods such as k-means need to get the number of clusters as an input parameter. Appropriate number of clusters may have an impact on clustering results. For example, the smaller number of clusters can help the clarity of the original data structure but the hidden information cannot be detected. On the other hand, using large number of clusters may result in high heterogeneity of the same cluster sets but it ignores the basic structure of data [ 3] . In recent years, due to their ability to solve various problems with very little change, evolutionary computation algorithms are widely used for clustering problems. On the other hand, these algorithms can manage the limitations in a good way. Hruschka et al. [ 9 ] have applied evolutionary algorithms to clustering problems in which the genetic algorithm is used employed. In this study, a simple coding procedure manipulates the fixed length individuals. The objective function maximizes the between-cluster similarity and between-cluster dissimilarity. In addition, this procedure finds the true number of clusters based on external criteria. There are various kinds of evolutionary methods that have studied clustering problems, such as evolutionary programming [ 10 ], particle swarm optimization [ 11 , 12 ], ant colony algorithm [ 13 ] or bee colony algorithm [ 14 ]. When the length of display solutions in evolutionary algorithms is equal to the number of data, then clustering efficiency will be reduced when there is a large number of data [ 15 ].So this paper suggests the dynamic clustering method based on discrete particle swarm optimization and genetic algorithm (DPSO-GA) . The proposed algorithm automatically determines the optimal number of clusters and simultaneously clusters the data set with minimal user intervention. The proposed method uses discrete PSO algorithm [ 16 ] and genetic algorithms with a new display solutions which are of variable length, we use the number of clusters and the primary centers of the clusters to specify the algorithm of our K-Means and thus raise the efficiency of clustering. The rest of the paper is organized as follows: In Section 2, particle swarm optimization algorithm and discrete particle swarm optimization algorithm is described. In Section 3 and Section 4, the genetic algorithms are briefly explained. In Section 5, the details of the proposed algorithm are given. Experimental results on five databases are reported in space Section 6. Finally, concluding remarks are presented in Section 7. 2.Particle swarm optimization Algorithms (PSO) PSO Algorithms [17 ] acts on a population of particles. In PSO the practical solution is called a particle and a fitness value of each particle is given through the objective function. Each particle has its own velocity (V id) and location (X id). The velocity and position of particle is modified according to its own flying experience and the population of particles. Thus, each particle moves towards the probability of best given ( pbest ) and the global best ( gbest ). Initially, the algorithm randomly produces the initial May-June 2056 ISSN 1013-5316; CODEN: SINTE 8 velocity (V id) and the initial position (X id Points) of each particle. Each particle remembers its best fitness value, and further, the individual values are compared to determine the best of the best called global best ( gbest ). Finally, the velocity and position of each particle is modified by the optimal individualal solution and the global best solution. The optimal solution is determined by repeating the calculations. Therefore, the update equation of the particle velocity and position are the relations (1 ) and (2). 𝑣𝑖𝑑 = 𝑤𝑣𝑖𝑑 + 𝑐1 × 𝑟𝑎𝑛𝑑1 × (𝑝𝑏𝑒𝑠𝑡𝑖𝑑 − 𝑋𝑖𝑑 ) + 𝑐2 × 𝑟𝑎𝑛𝑑2 × (𝑔𝑏𝑒𝑠𝑡𝑑 − 𝑋𝑖𝑑 ) (1) Xid = X id + vid (2) Where c 1 and c 2 are the individualized and global learning factors respectively and rand 1 and rand 2 are random variables between 0 and 1, and w 1is the inertia factor. 3.Discrete PSO Algorithm Discrete PSO algorithm deals with discrete variables and the length of solutions (particles) are according to the number of data (p = number of data). Amount of each component of the particle (vector) is between 1 and n (k = number of clusters) that represent the clusters that have been assigned to it. Particle X i records the best acquired place into a separate particle called pbest i. The Particle populations, in addition, save the best global position demonstrated by gbest [ 16 ]. The initial population of particles is produced by a set of randomly generated integers between 1 and n. The particles are fixed-length strings of integers are encoded as follows: Xi = {xi1 , xi2 , … , xip } in which xij ∈ {1,2, . . , n}, i = 1,2, . . , m , j = 1,2, … , p. For example, given the number of 3 clusters (k = 3) and 10 data (p = 10), the particle X1 = {1,2,1,1,2,2,1,3,3,2} indicates a candidate solution that in this way, data solutions 1, 3, 4, 7, have been assigned to Cluster 1and data 2, 5, 6 and 10 to Cluster 2 and data 8 and 9 to Cluster 3. The particles population thus is produced and fitting of each particle is calculated. The particle velocity is a set of transfers applied to solutions. The rate of (transfer) each particle is calculated by subtracting the two positions (of particles). The difference between X i and pbesti suggest changes that will be necessary for particle i to move from X to pbest. μ is the number of elements against zero in subtraction. If the difference between an element of X i and pbesti is zero, it means that there is the possibility to change the location and that element is subjected to change by the operation described below. First, the new vector P is generated in which the locations of elements in subtracting vectors against zero are stored. We generate a random number β. This number (β ) corresponds to the number of changes that must be made on the basis of the difference between X i and X i pbest i , so β lies in the range (0, μ ). Then we generate the binary vector Ψ according to measure μ where each component is associated with a component of vector P. The β number of elements in the vectorΨ is randomly initialized with 1 and the remainder is equal to zero. If the binary number is equal to 1 means that the component must change, and if the binary number is equal to zero, there is no need to change . A similar process is run based on the updated particle position and the Sci.Int.(Lahore),27(3),2055-2061,2015 global best position to calculate the particle in next step [16]. For instance, if (𝑋𝑖 − 𝑝𝑏𝑒𝑠𝑡𝑖 ) = (1 − 1, 2 − 3, 1 − 2, 1 − 1, 2 − 2, 2 − 1, 1 − 2, 3 − 3, 3 − 2, 2 − 1) = (0, −1, −1, 0, 0, 1, −1, 0, 1, 1) (3) The new P vector is equal to P = (2,3,6,7,9,10) and suppose μ = 6 , β = 5 and ψ = (0,1,1,1,1,1). This means that the element positions 3, 6, 7, 9 and 10 in the vector X i are initialized by the same element positions from pbest i. vector. The process is repeated with new particle and gbest until the final particle is achieved. 4.Genetic Algorithms GA algorithm is based on the idea of "survival of the fittest in natural selection" suggested by Darwin. Algorithms GA consists of selection operators 2 , reproduction operators 3 crossover 4 and mutation operators 5 . The fitted genes passed on the offspring by the best parents of the generation to reach the optimal solutions during the search relatively quickly. GAs is employed in many applications. 5.The proposed algorithm In the proposed method, a new display is used to show the particles in order to reduce the size of particles. In the following sections, first the new display and crossover, mutation and elitism operators will be described and then the proposed algorithm is given. 5.1 Display In the proposed method for clustering procedure the coding method by Falkenaure [ 18 , 19 ] in combination with the aforementioned coding method [ 15 ] are used.. This means that every individual has a variable length for displaying solutions. In the proposed display each individual is composed of two parts c = [l, g] in which the first part refers to element section and the second part shows the grouping section of the individual. In element section, every position is assigned to Element in every individual, every place into a small group of data (sub X). In this method, the stored value in this position shows the cluster to which the data in the X group belong. For example, the different parts of the solution (individual) for clustering of N subgroups and k clusters is shown by following relationship (4 ). l1 , l2 , … , lN |g1 , g 2 , … , g k (4) li Indicates a cluster that the data in the subgroup i have been assigned to it, while the group section represents the list of tags associated with solution clusters. lj = g i ⇔ ∀xk ∈ Xj , xk ∈ Gi (5) Note that the length of the element for the given problem is fixed (equal to N), but it is not constant throughout the length of the group and will vary from individual to individual. Therefore, the proposed method does not require the number of clusters as an input parameter but investigates the fitted value of k based on the objective function. For example, Figure 1 shows a model of proposed method solution to the problem of clustering. In this problem, of data clustering lies within five subgroups so the length of the element equals to 5.In this solution, the subgroup data X 1 belongs to cluster G1 , the subgroup data X 2 and May-June Sci.Int.(Lahore),27(3),2055-2061,2015 ISSN 1013-5316; CODEN: SINTE 8 Figure 1 Particle display in the algorithm DPSO-GA. X 5 belongs to G2 and the subgroup data X 3 and X 4 belongs to cluster G3 . With this display methodology, the near data must first be grouped into Xi subgroups. The closest neighbor for each data, k methodology can be used to use to get the groups X i. 5.2 Crossover Operator Crossover operator used in genetic algorithm of this paper is similar to [ 1 ] which has been accepted for the clustering problem. This crossover operator creates a child from two parents composed of the following steps: Initially, two parents are chosen randomly and crossovering is done on two points of their group. Elements related to selected groups of individuals within the child come first. Elements from the selected group of second individual enter a second child if the elements are not initialized previously by the first individual. Elements that have not been initialized before are assigned to the current groups. Empty clusters are removed. Current group labels in children are modified from 1 to k. Relationship (6 ) is another example of the crossover operator implemented in this paper. ind1 = [1 3 2 1 4 1 1 2 3 2 1 3 4 2 1|1↓ 2 3↓ 4] a) ind2 = [3 1 2 1 3 2 2 1 3 1 2 3 2 2 2|↑ 1 2↑ 3] b)off = [− 3 2 − − − −2 3 2 − 3 2 3|2 3] c)off = [− 3 2 1́ − 2́ 2́ 2 3 2 2́ 3 − 2 2́|2 3 1́ 2́] d)off = [3 3 2 1́ 1́ 2́ 2́ 2 3 2 2́ 3 2 2 2́|2 3 1́ 2́] e)off = [2 2 1 3 3 4 4 1 2 1 4 2 1 1 4|1 2 3 4] (6) 5.3 Mutation Operator Mutation operator makes small modifications in each individual of population with a low probability of occurring in A and D to discover the new areas of the search and when the algorithm is close to the convergence escapes the local optimum [1]. 1-Mutation by dividing cluster: dividing the selected cluster into two different clusters. Samples which belong to the main cluster with equal probability are assigned to the new cluster. Note that one of the new clusters produced keeps its class tag in the group section while new label of k + 1 is assigned to the other group. 2057 Selecting the first cluster to divide the cluster size is associated with larger clusters are more likely to split. For example, the application of this operator on the individual used in the mutation operator of relation (6 ) is shown in equation (7 ) in which cluster 1 is selected to divide. 22 1 3 3 4 4 5 2 1 4 2 5 1 4|1 2 3 4 5 (7) Mutation operator and merging: Merging two existing clusters selected randomly is done within a cluster. As for the mutation in cluster division, the probability of cluster selection depends on the size of clusters. To show the function of this operator, this operator is implemented on the individual whom the mutation operator is performed on (relation (6 )). In this case, suppose two clusters 2 and 4 have been selected for merging. 2 2 1 3 3 2 2 1 2 1 2 2 1 1 2| 1 2 3 (8) Note that the two mutation operators are applied in a sequence (one after another) with independent probabilities. Comparative form of the mutation operator probability is implemented. In this case, the mutation probability in the first generations is lower and get higher in the last generations to have the opportunity to escape from the local optimum. j Pm (j) = Pmi + (Pmf − Pmi ) (9) TG Pm (j) is the implemented mutation probability the j generation, TG indicates the total number of generations and Pmi and Pmf show the initial and final values, respectively. 5.4 Replacement and elitism In the proposed approach, the elitist approach has been used for the construction of the next generation. In elitism, the best individuals in j generation automatically are placed to the j + 1 generation to ensure that the best individuals found so far remains in the evolution by the algorithm 5.5 Proposed algorithm The algorithm proposed in this paper uses a discrete PSO algorithm and genetic algorithm in which the mutation and genetic crossover operators are added by discrete PSO method. It is expected that the combined algorithm increase the global search ability and escape from local optimal solution. Discrete PSO algorithm which has memory can establish the rapid correlations based on current optimal solution with the velocity of the rapid convergence. But when the optimal particle lens is a local optimal solution, the global optimal solution cannot be found. Therefore, this paper uses genetics to overcome this drawback. So, a parent is produced by PSO and then the parent uses the mutation and crossover operator in genetics to produce another parent. Finally, the next generation of parents selected by elitism is determined. This selection process continues until the terminating condition is met. Then k-means is used to modify the cluster centers. Figure 3 shows the flowchart of the proposed algorithm. May-June 2058 ISSN 1013-5316; CODEN: SINTE 8 Start Subgroups formation: The subgroup to which each data belongs is calculated by use of nearest neighbors. Generation of initial population Determine the best overall solution (gbest) and the best solution of each particle (Pbesti) Update the position of each particle (Solution) Update the population 1 Generation of population 1 Update the population 2 Crossover between the best of each particle (gbest) and the general (pbest) Mutation of the general best (gbest) Generation of population 2 Elitism selection Is the end condition satisfied? K-means algorithm End Figure 2 Flowchart of the proposed algorithm. May-June Sci.Int.(Lahore),27(3),2055-2061,2015 Sci.Int.(Lahore),27(3),2055-2061,2015 ISSN 1013-5316; CODEN: SINTE 8 At first, the algorithm k is used to define the subgroup nearest neighbors of each data to which they belong. Then, the particle population is randomly generated based on the display method mentioned in section 5.1. The length of each particle is equal to the total number of subgroups where data is located and the number of clusters that the subgroups were assigned to, the length of the particles is changing. Then, a discrete PSO and genetic algorithms are used to find the initial cluster centers and the number of clusters after a specified number of iterations. Then K-means is used for setting more precise (better) results. The steps of the proposed algorithm are as follows: Step 1- Parameters values include population size (number of particles), the maximum number of clusters (N c), and the mutation rate are initialized. Step 2- The subgroups of data using algorithm k, the nearest neighbors are determined. Step 3- The initial position (X i) of a single particle in a population is generated according to equation (4 ). Step 4: The values of the fitness of all particles are calculated. Step 5- Pbest i (the probability of the best given particle i so far) and gbest (the global best solution found so far) is specified. Step 6: The place of each particle is updated according to the procedures set forth in Section 3. Step 7- Perform steps 1 and 2 below on the parent updated in Step 6: Copy all the particles to generate the population 1. Do Crossovering on two points calculated in step 5 on Pbest i and gbest and the mutation on gbest. The population 2 is produced by this process. Step 8- Populations 1 and 2 are combined and the fitness value of each particle is calculated. Step 9-Elitism selection is applied on populations 1 and 2 to generate the next generation Step 10- Return to step 4 until a predefined number of iterations will be obtained. Step 11- According to gbest, the initial cluster centers and the number of clusters are identified and then the identified parameters are given to k-means algorithm as an input to group the data into clusters. 6. EXPERIMENTS Five data sets are used for clustering analysis and the results of the proposed algorithm in comparison with DCPG algorithms [ 3 ] is presented according to these 5 databases. 6.1 Databases In this paper, different methods are tested on 5 databases from UCI database. The database properties are provided in the table . Table 1 Database Properties Number of Number of Database data Properties Iris 150 4 Wine 178 13 Number of Classes 3 3 Glass 214 9 6 Seed 210 7 3 Tissue 106 9 6 2059 6.2 Preprocessing Data preprocessing is done on the basis that the maximum and minimum values for each attribute in the data set is found and then each attribute is subtracted from its minimum value and the obtained value is divided by the difference between maximum and minimum value of that attribute. [3]. Equation (10) described methods to obtain a normalized value of attribute i of data x: x −x xi = i min (10) xmax −xmin 6.3 Experiments Results and the Analysis Setting parameters for the algorithm are shown in Table 2: Table 2 Setting the values of algorithm parameters DPSO-GA Value Description of parameters 500 The total number of iterations 0.9- 0.1 Minimum and maximum values of the mutation rate 20 Population size 1 crossover rates In addition to the parameters listed in Table 1 , the maximum number of clusters should be identified for the implementation of the algorithm. Zhang et al. [ 20 ] stated that the number of maximum clusters should not be more than the root of the number of data in the dataset ; so the maximum values of clusters for iris and Wine databases equals to 13, for the Glass database is 15, Tissue database equals to 11 and for the seed database is 15. The number of experiments for each database is equal to 20 and the result average of these 20 times is reported as the results of the algorithm. 6.4 Estimation Algorithm Clustering Quality measurement method is based on the fact that how well-connected the things are to each other. Former clusterings stressed high similarity clustering of data set in the same cluster and attempted to find the shortest distance between data and the cluster center. In dynamic clustering, the distance between and within clusters are considered as measurement index. As a result, the modified index Turi [ 21 ] is accepted in relation (11 ) . intra VI = (c × N(0,1) + 1) × (11) inter Where (c × N(0,1) + 1) is considered as penalty term to prevent too many clusters; c is a constant equals to 30 and N(0,1) shows a Gaussian function with the mean of 0 and the standard deviation of 1 from the number of clusters. N(μ, σ) = 1 √2πσ e 2 (k−μ)2 ] 2σ2 [− (12) Equation (12 ) shows the Gaussian distribution function with the mean of µ and the standard deviation of σ. Turi shows that the result of dynamic clustering is placed in range 2 of the maximum number of clusters. Penalty term can prevent the data to group in two clusters and enable the clustering algorithms to find the appropriate number of clusters [ 21 ] . In this experiment, since Seed , Iris and Wine have a small number of clusters, N (0.1) is accepted to calculate the VI index and since Glass and Tissue have more clusters N (2,1) is admitted instead of N (0, 1) to calculate the VI index May-June 2060 ISSN 1013-5316; CODEN: SINTE 8 Np Where Np represents the total number of data. Smaller value of intra indicates better performance of clustering for algorithms. Finally, inter is the distance between two clusters defined by the equation (14 ) . inter = min{‖mk − mkk ‖2 } (14) ∀k = 1,2, . . , K − 1 , kk = k + 1, . . , K Inter focuses on the minimum distance between clusters and is defined as the minimum distance between cluster centers. Another criterion used to evaluate clustering is the Rand index [ 22 ] calculated the similarity between the obtained partition and the optimal solution, it means the percentage of right decisions taken by the algorithm. TP+TN R(U) = (15) TP+FP+TN+FN Where TP and TF are, respectively, the number of correct or incorrect attributions. When the decision involves assigning two elements to the same cluster. TN and FN show the number of correct and incorrect attributions when the decision involves assigning two elements to different clusters. Note that R is in the range [ 1,0 ] and the R-value is closer to 1 indicates better quality of solutions . 6.5 EVALUATION OF RESULTS Clustering performance indicators are the VI measured values , if the VI is smaller (larger) the results of clustering algorithm will be better (worse). VI value shows that the training process obtains convergence algorithm. Considering the number of clusters, the obtained number of clusters by the algorithm nearer to the correct number of clusters shows the better results of clustering algorithm. In the experiment, some comparisons are also made on the Rand-index values and the number of clusters. Since the initial solution is generated randomly, evaluation is done based on the mean values of 20 times experiments and standard deviations. The results based on the Rand index, the number of clusters and the standard deviation from the implementation of various algorithms on five databases are reported in Tables (3)- (7). As a result, the proposed method achieved better performance in terms of determining the number of clusters compared to other methods. Table 3 Comparison of results obtained by different methods on Iris Database Rand 0:01 ± 0.8 7 12:03 ± 0.86 Number of clusters Methods 0.3 ± 3.1 DPSO-GA 0.4 ± 2.95 DGPSO Table 4 Comparison of results obtained by different methods on Glass Database Rand Number of clusters Methods 0:01 ± 0.66 0.6 ± 5.45 DPSO-GA .02 ± 0.63 0.0 ± 5 DGPSO Sci.Int.(Lahore),27(3),2055-2061,2015 Table 5 Comparison of results obtained by different methods on Seed Database Rand Number of clusters Methods 0.0 5 ± 0.84 0.48 ± 2.85 DPSO-GA 0.05 ±0.75 0.37 ±2.15 DGPSO Table 6 Comparison of results obtained by different methods on Tissue Database Rand Number of clusters Methods 0.02 ± 0.77 0.31 ± 5.1 DPSO-GA 0.08 ±0.70 0:44 ± 5.25 DGPSO Table 7 Comparison of results obtained by different methods on Wine Database Rand Number of clusters Methods 0.01 ± 0.93 0.22 ± 3.05 DPSO-GA 0.07 ±0.91 0.39 ± 2.95 DGPSO . The value of Intra is shown in relation (13 ) that is the average of intra-cluster distance. 1 intra = ∑Kk=1 ∑u∈Ck‖u − mk ‖2 (13) 7. CONCLUSION AND FUTURE WORKS This paper proposed the DPSO-GA algorithm to solve the problem of setting the number of clusters and also to find the appropriate number of clusters according to the data characteristics. Five UCI datasets are employed with different numbers of clusters, different sizes and different types of data to verify that the DPSO-GA algorithm can offer better clustering results. Using a discrete PSO and GA algorithm DPSO-GA achieved better results. Clustering is a data mining method, which focuses on identifying relationships between data. In future work, we plan to suggest a multi-objective method based on DPSO-GA to identify the relationships with multiple criteria. REFERENCES 1. S. Salcedo-Sanz , Et al. , "A new grouping genetic algorithm for clustering problems," Expert Systems with Applications, vol. 39, pp. 9695-9703, 2012 . 2. M. Omran , Et al. , "Dynamic clustering using particle swarm optimization with application in unsupervised image classification," in Fifth World Enformatika Conference (ICCI 2005), Prague, Czech Republic , 2005, pp. 199-204 . 3. R. Kuo , Et al. , "Integration of particle swarm optimization and genetic algorithm for dynamic clustering," Information Sciences , vol. 195, pp. 124140, 2012 . 4. D.-X. Chang , Et al. , "A genetic algorithm with gene rearrangement for K-means clustering," Pattern Recognition, vol. 42, pp. 1210-1222, 2009 . 5. T. Kanungo , Et al. , "An efficient k-means clustering algorithm: Analysis and implementation," Pattern Analysis and Machine Intelligence, IEEE Transactions on, vol. 24, pp.881-892, 2002 . 6. GJ McLachlan and KE Basford, "Mixture models. Inference and applications to clustering," Statistics: Textbooks and Monographs, New York: Dekker, 1988, vol. 1, 1988 . May-June Sci.Int.(Lahore),27(3),2055-2061,2015 7. 8. 9. 10. 11. 12. 13. 14. 15. ISSN 1013-5316; CODEN: SINTE 8 S. Dehuri , Et al. , "Genetic algorithms for multi-criterion classification and clustering in data mining," International Journal of Computing & Information Sciences, vol. 4, pp.143-154, 2006 . S. Mitra and H. Banka " Multi-objective evolutionary biclustering of gene expression data, " Pattern Recognition, vol. 39, pp. 2464-2477, 2006 . ER Hruschka and NF Ebecken, "A genetic algorithm for cluster analysis," Intelligent Data Analysis, vol. 7, pp. 15-25, 2003 . M. Sarkar , Et al. , "A clustering algorithm using an evolutionary programming-based approach," Pattern Recognition Letters, vol. 18, pp. 975-986, 1997 . T. Cura, "A particle swarm optimization approach to clustering," Expert Systems with Applications, vol. 39, pp. 1582-1588, 2012 . S. Das , Et al. , "Automatic kernel clustering with a multi-elitist particle swarm optimization algorithm," Pattern Recognition Letters, vol. 29, pp. 688-699, 2008 . H. Jiang , Et al. , "Ant clustering algorithm with Kharmonic means clustering," Expert Systems with Applications, vol. 37, pp. 8679-8684, 2010 . C. Zhang , Et al. , "An artificial bee colony approach for clustering," Expert Systems with Applications, vol. 37, pp. 4761-4767, 2010 . L. Cagnina , Etal. , "An efficient Particle Swarm Optimization approach to cluster short texts," Information Sciences, 2013 . 16. 17. 18. 19. 20. 21. 22. May-June 2061 H. Sahu , Et al. , "A discrete particle swarm optimization approach for classification of Indian coal seams with respect to their spontaneous combustion susceptibility," Fuel Processing Technology, vol. 92, pp. 479-485, 2011 . RC Eberhart and J. Kennedy, "A new optimizer using particle swarm theory," in Proceedings of the sixth international symposium on micro machine and human science 1 995 , pp. 39-43 . E. Falkenauer, "The grouping genetic algorithmswidening the scope of the GAs," Belgian Journal of Operations Research, Statistics and Computer Science, vol. 33, p. 2, 1992 . E. Falkenauer, Genetic algorithms and grouping problems : John Wiley & Sons, Inc., 1998 . D. Zhang , Et al. , "A dynamic clustering algorithm based on PSO and its application in fuzzy identification," in Intelligent Information Hiding and Multimedia Signal Processing, 2006. IIHMSP'06. International Conference on , 2006, pp. 232235 . RH Turi, Clustering-based colour image segmentation : Monash University PhD thesis, 2001 . WM Rand, "Objective criteria for the evaluation of clustering methods," Journal of the American Statistical association, vol. 66, pp. 846-850, 1971.