Вы находитесь на странице: 1из 7

JOURNAL OF COMPUTING, VOLUME 4, ISSUE 5, MAY 2012, ISSN 2151-9617 https://sites.google.com/site/journalofcomputing WWW.JOURNALOFCOMPUTING.

ORG

157

Combined use of Community Detection and Particle Swarm Optimization to Support Decision Making
Nesrine Ben Yahia*, Narjs Bellamine*, and Henda Ben Ghsala*
AbstractDecision making whenever and wherever it is happening is crucial to organizations success. In this paper, we suggest a multimodal decision making (MDM) process where decision may be made by three possible modes: individual, collective or hybrid. In order to support multimodal DM process, we combine the use of a graph clustering technique i.e. community detection and a meta-heuristic optimization technique i.e. particle swarm optimization (PSO). At the beginning, a community detection thresholding algorithm based on both structural and attribute clustering is advanced to compute nodes and edges weights. In order to measure the performance of clustering, we use network modularity, a metric that considers cohesion and coupling of communities. Then, PSO is employed to maximize network modularity by finding out the threshold that leads to cluster the network into two communities with high cohesion and weak coupling. Finally, we evaluate the proposed algorithm with an example of application. Index TermsDecision making, multimodal process, community detection, particle swarm optimization.

1 INTRODUCTION
propose to structure and formalize interaction between participants. Then, [6] propose a role of facilitator to manage and organize the different process tasks. [8] propose also ontology to structure the problem which facilitates the find out of similar problems in the organizational memory and then to look at the different solutions that are generated to solve the detected similar problem.

In this paper, we show how it is interesting to use community detection techniques to support decision making process. Community detection consists on clustering networked systems such as social networks and computer networks into communities with high cohesion and weak coupling. For example, community detection is presented in [3] as a class of new algorithms for social networks clustering i.e. the discovery of community structure in social networks. The aim with community detection in this paper is to identify appropriate participants within the same organization that can be invited to help in the decision making process. Thus, we aim to cluster the organization (represented by a social network, where nodes represent participants and edges represent relationships between participants) into two communities with high cohesion and weak coupling basing on both structural (considering the edges weights) and attribute (considering the nodes weights) clustering. Then, to measure the performance and the quality of clustering, we use the metric modularity introduced in [3] as a measure of the quality of a particular division of a network that considers cohesion and coupling of communities. Next, Particle Swarm Optimization (PSO), a meta-heuristic optimization technique, is employed to maximize network modularity. This paper proposes and explicates the multimodal de cision making (MDM) process at three levels (individual, RIADI-GDL laboratory, National School of Computer Sciences, University collective and hybrid) and it aims to explicit the support of Manouba, Tunisia. of the MDM process by community detection algorithm. Then, we use particle swarm optimization to optimize the

ECISION making (DM) whenever and wherever it is happening is crucial to organizations success. DM represents process of problem resolution based on four phases as it is proposed by [1] and revisited by [2]: intelligence, design, choice and review. In the intelligence phase, the problem is identified. In the design phase, the proposed alternatives or solutions are generated. In the choice phase a solution is selected. Finally, in the revision phase the choice is revised and an intelligent feedback permits to correct errors. In this paper we propose to support DM process using community detection defined in [3] as a graph clustering technique. Decision making process is largely defined as a collection of correlated tasks including gathering and exchanging information, identifying problem, choosing among alternatives and implementing a choice [4-5]. It is also defined as a multi-participant process characterized by a high level of interactivity where we move from an individual toward a group activity [6] and it is subdivided into three phases: pre-meeting, during the meeting and post meeting. In the pre-meeting, the DM process is planned by defining an agenda of tasks and activities with time limits. Within during meeting phase, the generation and evaluation of alternatives and the solution choice are finalized. Finally, in the post- meeting phase, summary of meeting and reports are generated. In order to make correct decision, individuals and teams within organizations need support to make that more effective and efficient. There are different perspectives that are adopted to support DM: for example [7]

JOURNAL OF COMPUTING, VOLUME 4, ISSUE 5, MAY 2012, ISSN 2151-9617 https://sites.google.com/site/journalofcomputing WWW.JOURNALOFCOMPUTING.ORG

158

network modularity. The outline of this paper is as follows. In section two, we propose the multimodal decision making (MDM) process. In section three, we illustrate the crucial concepts involved in community detection and we present our community detection algorithm to support MDM process. In section four, PSO is used to optimize the proposed algorithm. Finally, section five presents an example of network where we will experiment the proposed method.

2 MULTIMODAL DECISION MAKING


Our aim in this section is to focus on multimodal decision making process. In this paper, specifically, we consider MDM process as a process of problem resolution based on four phases: intelligence, design, choice and review. a) In the intelligence phase, the problem is identified. b) In the design phase, the proposed alternatives or solutions are generated. c) In the choice phase a solution is selected. d) In the revision phase the choice is revised and an intelligent feedback permits to correct errors. The essential property of MDM here, we argue, is that it enables three modes of problem resolution and decision making: individual mode (computer human aspect where actors interact with their computer), collective mode (computer human-human aspect where actors interact together with their computers) and hybrid mode (navigation between the two previous cases). In this paper, we argue also that one of the interesting points to be considered in the MDM process is the actors dynamicity as it represents a process by which individuals formulate the problem (together), generate and evaluate solutions (together) and make decision (together). By actors dynamicity we mean it is not mandatory or surely to know beforehand all the actors who will be implicated in each phase of the process. An actor can be involved in one phase and doesnt participate in another phase. For example, an invited actor accepts to participate in the problem formalization phase but he will not propose solution so he will not participate in the solutions generation phases. Thus, we characterize MDM process by actor oriented perspective where actors guide the MDM process and orient the resolution mode in all phases (problem formulation, solutions generation or decision making). Dynamicity of actors is useful especially when we talk about hybrid mode resolution (when we navigate between the two modes individual and collective). For example, if one individual has a problem he can choose to construct and formalize it alone or with others to help him. After problem conception (individual or collective), the same individual can generate solutions alone or with others. After solutions conception (individual or collective), he can choose one alternative alone or with others. Accordingly, we distinguish between three types of actors: problem-constructor, solution-constructor and decision-maker: Problem-formulator: identify, formulate and structure the problem. His aim is to find out the problem. Solution-generator: generate and propose solutions.

His aim is to generate alternatives, set criteria and scenarios to evaluate alternative. Decision-maker: select and choose one alternative. His aim is to choose alternative(s) and determine the outcome of chosen alternatives. If we consider A as the set of actors participating to the different activities of the MDM so A=<Ap, As, Ad> where Ap represents the actors that participate to the problem conception and |Ap| = nbr(problemformulators), As represents the actors that participate to the solutions conception and |As| = nbr(solutiongenerators), Ad represents the actors that participate to the decision making, |Ad| = nbr(decision-makers). In the problem conception, MDM process can be characterized as single problem- formulator or multi problemformulator. In the solutions conception, MDM process can be characterized as single solution- generator or multi solution- generator. In the selection, MDM process can be characterized as single decision-maker or multi decisionmaker. Thus, the three modes of resoulution are described as follows. Individual mode of problem resolution corresponds to the case single problem-formulator, single solutiongenerator and single decision-maker, so |Ap|= 1, |As|= 1and |Ad|= 1. Collective mode of problem resolution corresponds to the case multi problem-formulator, multi solutiongenerator and multi decision-maker, so |Ap|> 1, |As|> 1and |Ad|> 1. Hybrid mode of problem resolution corresponds to the rest cases.

3 COMMUNITY DETECTION SUPPORTED DECISION MAKING PROCESS


3.1 Overview of community detection In a graph, community represents a cohesive group of nodes that are more densely joined to each other than to the nodes in other communities [9]. Consequently, community consists on a group of nodes which are densely related to each other but sparsely related to other groups in the network or simply it is a set of nodes with better internal connectivity than external connectivity. For example, in a social network, a community can represent a group of friends and in the World Wide Web a community can represent a group of pages having same topics. There are two types of communities: overlapping and disjoined. Overlapping communities allow the share of nodes (each node can appear in more than one community) but disjoined communities avoid this (each node can appear in one community). Community detection is considered as a multidisciplinary technique where the use of graph partitioning, computer science, and sociology are combined. It aims to divide network nodes into groups within which the network connections are dense, but between which they are sparser [3]. It is also defined as typical applications of graph clustering in social networks [10] and so as social

JOURNAL OF COMPUTING, VOLUME 4, ISSUE 5, MAY 2012, ISSN 2151-9617 https://sites.google.com/site/journalofcomputing WWW.JOURNALOFCOMPUTING.ORG

159

network mining [11]. Approaches of community detection fall into two classes: agglomerative and divisive [12]. Agglomerative approaches are based on bottom-up processes that focus on the addition of edges to the network. They start with an empty network (n nodes with no edges) and consider the vertex pairs with highest similarity, calculate distances between nodes pairs, and then add edges to the network. For example [13] use this approach to discover communities in the network of collaborations between film actors where nodes represent actors, edge between two actors indicates that they have participated in the same film and edge weight measures how many films actors have appeared in together. Divisive approaches are based on top-down processes that focus on the removal of edges from the network. They start with a network of connected nodes and search the least similar connected nodes pairs and then remove the edges between them. By repeating this removal, the network is divided into smaller and smaller communities. For example [3] use this approach to discover communities in the network of interactions between characters in the novel Les Misrables of Victor Hugo where nodes represent characters, edge between two characters indicates that they have co-appeared together in the same scene and edge weight measures how many scenes characters have co-appeared in together. The differences between many community detection techniques fall into the definition of various criteria that are employed to denote the meaning of more densely groups during the network clustering. There are three different perspectives: Structure based clustering: where network division mainly focuses on nodes connectivity i.e. the topological structure of the network but it ignores the nodes properties which could be often heterogeneous. For example, [11] propose two topological algorithms CCS and CNS based solely on network structure. Attribute based clustering: where network division mainly focuses on attribute similarity i.e. nodes properties but it ignores nodes connectivity. For example, [14] propose SNAP algorithm that groups nodes with the same attribute values into one cluster. Structural/Attribute based clustering: where network division focuses on both structural and attribute similarities i.e. each partition contains densely connected nodes with homogeneous attribute values. For example, [10] propose SA-Cluster algorithm that is based on a unified distance to measure both structural and attribute similarities of nodes and [11] also propose NAS algorithm that based on node attributes and neighborhoods. These three techniques are experimented and evaluated in [10] using the DBLP Bibliography network, where nodes represent authors characterized by two attributes prolific (authors with < 10 papers are labeled as low prolific, authors with 10 and < 20 papers are labeled as prolific and authors with 20 papers are labeled as highly prolific) and primary topic (research topics of authors) then edges represent co-authorships (authors appear in the same paper).

To measure the performance of community detection techniques, there are different metrics that are proposed but the popular one is the metric modularity introduced in [3] as a measure of the quality of a particular division of a network. For a network with k communities, modularity is computed using a kk matrix where each element dij (ij) represents the fraction of edges between communities i and j and dii denotes the fraction of edges within community i. Modularity is then given by:
M = i (dii ( j dij)2) (1)

In social networks, Newman et al. [3] report that M generally falls between 0.3 and 0.7.

3.2 Community Detection supported multimodal DM Individuals and groups within organizations need support to make correct decision. In this paper, we attempt to achieve this support by identifying and grouping right people in the organization to help them during the decision making process. Here, we mean by right people all actors in the organization that are closely similar to the individuals or groups need the support with having interrelations with each other. To be closely similar in our case means to share maximum of attributes and we suppose that the connectivity of the actors in the organization is provided as part of the input. Consequently, an individual (respectively a group) who has a problem can invite others to help him/her in the MDM process. So the role here of the community detection is to provide to this individual (respectively to this group) list of appropriate participants in the same group or organization representing the right people that can be invited to help him in the DM process. Thus, the goal is to partition the organization into two disjoined clusters, one contains closely similar actors (with homogeneous attribute values) and the other contains the rest of the organization using agglomerative approach (we start with an empty network, compute the distances between all interconnected actors in the organization and detect the more densely community that contains individual/group that need support). So, we need a structural/attribute based clustering to detect community that can help in the MDM process. We suppose given a social network N= (A, R) where A contains (nodes) actors of organization participating to the different activities of the MDM process and R represents the relations (edges) between actors organization. a A, a = <a1, a2,, an> where ai represents attributes of a so the profile of a. We distinguish between two types of an attribute: nominal (as topics of interest) and continuous (with continuous values as the attribute age). For nominal ones we distinguish here between monovalued and multi-valued. Mono-valued attribute is an attribute with one possible value such as the nationality. Multi-valued attribute is an attribute with several possible values such as the preferences or topics of interest (so ai = <ai1, ai2,, aik> where aij represent possible values of ai). In this paper we adopt the NAS (Node Attribute Simi-

JOURNAL OF COMPUTING, VOLUME 4, ISSUE 5, MAY 2012, ISSN 2151-9617 https://sites.google.com/site/journalofcomputing WWW.JOURNALOFCOMPUTING.ORG

160

larity) community detection algorithm is introduced in [11] that considers both node attributes and neighborhoods where edges weightings are calculated as it is shown in the pseudo code of Fig. 1. Then, to detect communities in the network, after calculating and normalizing the edge weights to the range (0,1), they use a normalized thresholding method where any pair of nodes i and j whose edge weight exceeds the threshold are placed into the same community. To experiment this algorithm, [11] use the metric modularity given in (1) and proves that NAS algorithm values (over 0.91) exceeds the range (0.3, 0.7) reported by Newman et al. [3] thats why we attempt to reuse this algorithm.

node and the node representing the input profile of the community detection. The interest of nodes weights here is to calculate the distance between the actor that needs help and each actor in the same organization, For each edge between i and j, the weight W(i,j) is calculated by summing the similarity of attributes of node i and node j. In Fig. 2, we propose the function Calcul_weight( node i, node j) that permits to calculate both nodes weights and edges weights when similarities are detected for nominal (mono-valued or multi-valued) attributes or continuous attributes. Then, for nodes weights, the proposed function takes as parameters the node of input profile and each other node. Moreover, for edges weights, the proposed function takes as parameters all couple of nodes i and j that are interrelated. After calculating and normalizing nodes and edges weights, we use a threshold s where each node, that the weight exceeds s, is placed in the same community with the input node p. the pseudo code of DMS-NAS algorithm is defined in Fig. 3. Next, we can compute the modularity given by (1).

Fig. 1. NAS algorithm proposed in [11].

In this paper, we propose DMS-NAS (Decision Making Support- Node Attribute Similarity) which is inspired from NAS algorithm where the community detection input represents a profile and output represents the densely community containing the profile given in input. As for each phase of DM process, actor can be an individual (for individual phases) or a group (for collective phases), there are two possible inputs for the community detection algorithm: individual profile (sets of attributes characterizing the individual) and group profile (sets of attributes characterizing the group where attributes are generated by merging the nominal attributes and calculating the continuous ones).

Fig. 3. DMS-NAS algorithm pseudocode.

This algorithm can be used to detect community in three cases: to identify densely community to structure the problem and/or to identify densely community to generate solutions and/or to identify densely community to make decision.

4 PSO-BASED COMMUNITY DETECTION


In order to get more efficiency, we point to improve the previously proposed community detection algorithm by optimally adjusting the threshold s that maximizes the network modularity. Then, since the modularity is typically NP-hard to optimize exactly [15], we propose to use particle swarm optimization which consists on a metaheuristic technique for NP-hard optimization [16].

Fig. 2. Calculus of weights.

DMS-NAS algorithm is based on both nodes weights and edges weights. For each node j, the weight W(j) is calculated by summing the similarity of attributes of this

4.1 Overview of Particle Swarm Optimization Particle Swarm Optimization (PSO) is introduced by Kennedy and Eberhart in [17] as a new evolutionary computation technique inspired by social behavior simulation of bird flocking or fish schooling. We select this technique thanks to its ability to provide solutions efficiently with only minimal implementation effort and its fast convergence [18], PSO has attracted much attention

JOURNAL OF COMPUTING, VOLUME 4, ISSUE 5, MAY 2012, ISSN 2151-9617 https://sites.google.com/site/journalofcomputing WWW.JOURNALOFCOMPUTING.ORG

161

and been widely used in complex function optimization, neural network training and data mining. The population in PSO is called a swarm where the individuals, the particles, are candidate solutions to the optimization problem in the multidimensional search space (D dimensional). Each dimension of this space represents a parameter of the problem to be optimized. For each iteration t, every particle i is characterized by its position xi(t) and its velocity vi(t) which are usually updated synchronously in each iteration of the algorithm. A particle adjusts its velocity according to its own flight experience and the flight experience of other particles in the swarm in such a way that it accelerates towards positions that have had high objective function or the fitness function (the performance index of a population) values in previous iterations. There are two kinds of position towards which a particle is accelerated in common use. The first one, a particles personal best position achieved up to the current iteration, is called Pbesti. The other is the global best position obtained so far by all particles, called Gbest [19]. There have been several versions of Particle Swarm Optimization. We choose in this paper one of the basic versions as it focuses on cooperation rather than cooperation and is characterized by no selection. In this context, we choose the PSO version with constriction factor for its speed of convergence [20]. In this case, the modified position xi(t+1) and velocity vi( t+1) at the iteration t+1 for particle i can be calculated using the current velocity, Pbesti and Gbest as shown in the following formulas: xi(t+1) = xi(t) + vi(t) (2) Vi(t+1) = K * (vi(t) + c1 * r1* (Pbesti - xi(t+1)) + c2 * r2* (Gbest xi(t+1)) (3) Where K: the constriction factor given by K = 2/ | 2 - c c2 4c |With c=c1+c2 and c>4. r1 and r2: random numbers between 0 and 1. c1: self confidence factor. c2: swarm confidence factor. The evolution procedure of PSO Algorithm can be summed up in the following pseudo code 1. First of all, the problem dimension are fixed (PSO parameters), the positions and velocities initializations are defined. 2. The fitness of each particle is calculated. 3. Position and velocity of each particle are updated using (2) and (3). 4. If the maximum iteration number is reached, the final solutions are given. 5. Otherwise, we loop to 2 and so on.

with high cohesion and low coupling where the first will contain the appropriate participants to be involved in the decision making process and the second will contain the rest of the organization. So, in this paper, each particle in the swarm expresses a threshold that divides the organization into two clusters (one contains the nodes with weights that exceeds the threshold and the second contains the rest of nodes). For the present problem, we use network modularity as fitness function where its value is between 0 and 1and higher value represents better partition. For parameters initialization of PSO algorithm, we propose the following values: 1. Initial swarm population is composed of 10 particles, number of iterations is 1000 and we base our choice of c1 and c2 on the study done in [21] where c1= 2.8 and c2=1.3. 2. Initial positions and velocities of particles are generated by using random values in [Wnmin, Wnmax] where Wnmin and Wnmax are respectively the minimal and maximal values of nodes weights. 3. The initial fitness value of each particle xi is initialized to the modularity network corresponding to the threshold xi using (1). 4. The local best of each particle Pbesti is set to its current fitness value. 5. The global best value of the position Gbest is set to the best value of all Pbesti.

EXAMPLE

The aim in this section is to apply the proposed DMSNAS algorithm with an example based on an organization of 10 actors. Each actor is characterized by an ID and a set of attributes corresponding on his profile. As it is shown in Table 1, we consider here four attributes: Languages and Topics of interest which are nominal and multi-valued attributes, Nationality which is nominal and mono-valued attribute and Age which is continuous attribute). Then, we suppose given a graph representing the organization network where nodes represent the 10 actors (noted by their ID) and edges represent interactions or relations between actors (Fig. 4). We also suppose that actor profile characterized by the ID 1 will represent the input profile i.e. the actor who has the problem and needs to know right people in his organization to support him in the DM process. TABLE 1 EXAMPLE OF ACTORS PROFILES
ID 1 2 3 4 Actors Profiles
Languages Nationality Topics of interest Age

4.2 PSO-based community detection In order to optimize network modularity, many authors such as [16] use also PSO where each particle in the swarm expresses a partition of vertices in the network. So, PSO is used to determine the appropriate number of communities which leads to maximize the modularity. However, in our case we know this number because our aim here is to split the network into two communities

Arabic, English, French Arabic, French Arabic, French English, French

Tunisian Tunisian French French

CSCW, KM, DM IR, DB, KM AI, KM CSCW, BPM

25 30 28 42

JOURNAL OF COMPUTING, VOLUME 4, ISSUE 5, MAY 2012, ISSN 2151-9617 https://sites.google.com/site/journalofcomputing WWW.JOURNALOFCOMPUTING.ORG

162

ID 5 6 7 8 9 10

Actors Profiles
Languages Nationality Topics of interest Age

Arabic, English, Spanish Arabic, English, French English, French French, Spanish French, Spanish Arabic, English, French

Spanish Tunisian French Spanish French English

SOA, SMA KM, DM, KM DB, SGBD SE, OS WWW Sociology, CSCW

35 25 53 29 45 35

three modes of resolution (individual, collective and hybrid) and is characterized by dynamicity of actors where we separate between three types of actor (problemformulator, solution-generator and decision-maker). Then, we explain how it is useful to support MDM process by community detection. To identify right densely community that can help in the MDM, we propose a simple thresholding algorithm based on Structural/Attribute clustering where the threshold is automatically adjusted to maximize the network modularity using the meta heuristic Particle Swarm Optimization.

REFERENCES
[1] [2] H. Simon, 1977: The New science of management decision. Prentice hall, Englewood-Cliffs. P. Zarat, Des Systmes Interactifs dAide la Dcision Aux Systmes Coopratifs dAide la Dcision : Contributions conceptuelles et fonctionnelles., HDR dissertation, INP Toulouse, 2005. M. E. J. Newman, and M. Girvan, Finding and evaluating community structure in networks, journal of Physical Review E, Volume 69, Issue 2, 2004. Mintzberg H., The structuring of Organization, Prentice Hall, 1979. J.E. McGrath, Groups Interaction and Performance, Prentice Hall, Englewood Cliffs, NJ, 1984. A. Adla, J. L. Soubie, and P. Zarate, A cooperative intelligent decision support system for boilers combustion management based on a distributed architecture, Journal of Decision Systems. Volume 16, Number 2, 2007, pp. 241-263. M. Lewkowicz, Conception de collecticiels pour la gestion cooperative des connaissances, Ph.D. dissertation, Paris VI univ., 2000. O. Chourabi, M. Ben Ahmed, and Y. Pollet, Ontology Based Knowledge Management in System Engineering. Insight Journal of INCOSE, Volume 11, ISSUE 3, July 2008. M. A. Porter, J. P. Onnela, and P. J. Mucha,Communities in Networks, Notices of the AMS Volume 56, Number 9, October 2009, pp 1082 1097. Y. Zhou, H. Cheng, and J. Xu Yu, Graph Clustering Based on Structural/Attribute Similarities, VLDB09, August 2428, 2009, Lyon, France. K. Steinhaeuser, and N. V. Chawla, Community Detection in a Large Real-World Social Network International Conference on Social Computing, Behavioral Modeling and Prediction, Phoenix, Arizona, USA, 2008, pp. 168-175. J. Scott, Social Network Analysis: A Handbook. Sage Publications, London, 2nd edition (2000). L. A. N. Amaral, A. Scala, M. Barthlmy, and H. E. Stanley, Classes of small-world networks. Proc. Natl. Acad. Sci., vol. 97, no. 21, 2000. Y. Tian, R. A. Hankins, and J. M. Patel. Effcient aggregation for graph summarization. In Proc. 2008 ACM-SIGMOD Int. Conf. Management of Data (SIGMOD08), Vancouver, Canada, 2008, pp. 567580. Jure Leskovec, Kevin J. Lang, Michael W. Mahoney, Empirical Comparison of Algorithms for Network Community Detection WWW 2010, April 2630, 2010, Raleigh, North Carolina, USA. ACM 978-1-60558799-8/10/04. M. Clerc, P. Siarry. Une nouvelle mtaheuristique pour loptimisation difficile: la mthode des essaims particulaires, Journal sur l'enseignement des sciences et technologies de l'information et des systmes, Vol. 3, n. 7, 2004. R. C. Eberhart and J. Kennedy, " New optimizer using particle swarm theory," in Proceedings of the 6th International Symposium on Micro Machine and Human Science, pp. 3943, Nagoya, Japan, October 1995.

[3]

Fig. 4. Graph representing the network organization.

[4] [5] [6]

We use now DMS-NAS algorithm to compute nodes and edges weights. For the nominal attributes (Languages, Topics of interest and Nationality) the weight is incremented by 1 for each similar attributes but for the continuous attribute (Age), we set to 0.035 ( =1/28 where 28 is the difference between the minor age 25 and the greater one 53). Fig. 5 shows the results of the nodes and edges weights calculus. Then, we use the PSO algorithm to find out the threshold s that will maximize the network community. All nodes which weight exceeds s are considered in the output community (with green color).

[7] [8]

[9]

[10] [11]

[12] [13] [14] Fig. 5. Output community contains green nodes. [15]

So, we divide the network into 2 communities, the first one contains the nodes 1,2,3,6,10 which represent the right people to help and participate in the MDM process and the second one contains the rest of nodes. In this example, the threshold s is 0.4 and the modularity M is 0.71.

[16]

6 CONCLUSION
In this paper, we introduce Multimodal Decision Making process as a problem resolution process that covers

[17]

JOURNAL OF COMPUTING, VOLUME 4, ISSUE 5, MAY 2012, ISSN 2151-9617 https://sites.google.com/site/journalofcomputing WWW.JOURNALOFCOMPUTING.ORG

163

[18] Parsopoulos K. E and Vrahatis M. N, 2010. Particle Swarm Optimization and Intelligence: Advances and Applications. Information Science Reference (an imprint of IGI Global), United States of America. [19] J. Yisu, J. Knowles, L. Hongmei, L. Yizeng and D.B. Kell, 2008. The Landscape Adaptive Particle Swarm Optimizer. Applied Soft Computing, 8, pp.295- 304. [20] M. Clerc, 1999. The Swarm and the Queen: Towards A Deterministic and Adaptive Particle Swarm Optimization. Proceedings of the Congress of Evolutionary Computation, Washington, Dc, pp. 19511957. [21] A. Carlisle and G. Dozier, 2001. An Off-The-Shelf PSO. Proceedings of the Particle Swarm Optimization Workshop, Indianapolis, Ind, Usa, pp. 16. [22] Z. Shi, Y. Liu, J. Liang. PSO-based Community Detection in Complex Networks. Second International Symposium on Knowledge Acquisition and Modeling, 2009, pp. 114-119.

Nesrine BEN YAHIA is currently preparing her PhD thesis and teaching at the National School of Computer Sciences (Tunisia). She studied computer sciences at National School of Computer Sciences (Tunisia) where she obtained the Engineering degree in 2008 and the Master degree in 2009. Her research interests include decision making, computer supported collaborative work and knowledge management. Dr. Narjs BELLAMINE is an associate professor of computer science at the University of Manouba, Tunisia, and a researcher in RIADI-GDL laboratory of the National School of Computer Sciences. Dr. Bellamine received her engineering diploma and masters degree from ENSEEIHT Toulouse and her PhD from University of Toulouse 1, France. Her main research include complexity theory, agent-based modeling and simulation of cooperative systems, computer supported collaborative work and groupware development. Prof. Henda BEN GHESALA is a Professor of Computer Science at the University of Manouba, Tunisia. She is Director of the RIADI -GDL laboratory of the National School of Computer Sciences. Her research interests lie in the areas of information modeling, databases, temporal data modeling, object-oriented analysis and design, requirements engineering and specially method engineering.

Вам также может понравиться