Semantic and Structural Performer Clustering in BPMN Models

Transformed into Social Network Models

Wiem Khlif

and Hanêne Ben-Abdallah

1,2

Mir@cl Laboratory, University of Sfax, Sfax, Tunisia

King Abdulaziz University, Jeddah, K.S.A.

Keywords: BPM, Knowledge Discovery, Knowledge Rediscovery, Affiliation, Restructuring, Social Network Model,

Hierarchical Clustering.

Abstract: Current trends in organization restructuring focus on the social relationships among the organizational

actors in order to improve the business process. Proposed business process model restructuring approaches

adopt either social network discovery or rediscovery techniques. Social network discovery uses semantic

information to guide the affiliation process during its analyses, whereas social network rediscovery uses

structural information to identify groups in the social network. In this paper, we propose a hybrid method

that exploits both knowledge discovery and rediscovery to suggest a new structure of a business process

model that is based on performers clustering. Using the context concept, the proposed method applies a

hierarchical clustering algorithm to determine the performer partitions; the algorithm uses two newly

defined distances that account for the semantic and structural information. The method is illustrated and

evaluated experimentally to analyze its performance.

1 INTRODUCTION

Among the recent research efforts to making

Business Process Management (BPM) more

efficient, several researchers have been investigating

restructuring techniques that are centered on the

organizational perspective or people (Oinas-

Kukkonen et al., 2010). The main hypothesis of

these techniques is that social relationships among

people or organizational units affect the overall

performance of the business process model. Starting

from this hypothesis, several researchers have been

examining how to apply the concept of social

network and its analysis methods to business process

modelling. Their objective is to restructure the

organization so that its business process model

becomes more "efficient".

The so-far proposed approaches adopting social

network techniques for BPM can be divided into two

categories: social network rediscovery (Van der

Aalst et al., 2005) (Choi et al., 2007), (Song and Van

der Aalst, 2008) (Hong et al., 2012) (Boulmakoul

and Besri, 2013), and social network discovery

(Battsetseg et al., 2013) (Kim, 2013). Social network

rediscovery-based approaches extract structural

information from the business process event logs to

identify the connections among the performers or

organizational units, e.g., work transfers (Hong et

al., 2012).

In contrast, social network discovery-based

approaches explore the semantic perspective of a

business process model (e.g., the performers' roles)

to identify the social relationships among

organizational performers and units. Certainly, both

the structural and semantic information within an

organization are correlated and influence one

another. Hence, using exclusively either a

rediscovery approach or a discovery approach

reduces the scope of possible analyses that can be

made. Consequently, this may reduce the domain of

possible restructuring solutions.

Our objective in this paper is to use both the

knowledge discovery and rediscovery approaches to

find an affiliation of well-connected performers (the

structural aspect) that have similar profiles (the

semantic aspect). To do so, we introduce a new

definition of affiliation that includes both aspects,

and a new community detection method based on

the new definition. The community detection

method uses two new distances we define to account

for the structural and semantic aspects. It is based on

a hierarchical clustering algorithm that partitions the

performers (actors and/or organizational units) into

Khlif W. and Ben-abdallah H..

Semantic and Structural Performer Clustering in BPMN Models Transformed into Social Network Models.

DOI: 10.5220/0005555800790086

In Proceedings of the 10th International Conference on Software Engineering and Applications (ICSOFT-EA-2015), pages 79-86

ISBN: 978-989-758-114-4

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

sets of well-connected performers with similar

profiles. The connection reflects the structural/work

flow dependencies among the performers within the

organization, whereas the profile similarities reflect

the semantic relationships among them--e.g., in

terms of their affiliation to a pool and to a lane, their

assigned roles, and their permissions to perform the

activities.

Once the performer communities are identified,

we can apply the set of graph optimization rules we

proposed in (Khlif and Ben-Abdallah, 2015). These

rules combine the semantic and structural aspects to

reduce the control flow complexity of a business

process modelled in the Business Process Modelling

Notation (ISO/IEC 19510, 2013).

The remainder of this paper is structured as

follows: Section 2 overviews existing approaches for

organization restructuring. Section 3 presents the

definition of the context concept. Section 4 shows

how the new context concept can be used in a

method for identifying performer affiliations. In

section 5, we summarize the presented work and

outline its extensions.

2 RELATED WORK

2.1 Rediscovery-based Approach

Adopting a rediscovery approach, (Boulmakoul and

Besri, 2013) combine structural analysis with Q-

analysis and Social Network Analysis (SNA)

techniques. SNA plays an important role since it

evaluates the relationships among performers, roles,

units and even an entire organization (Stanley and

Katherine, 1999). This kind of analysis can extract

important information to improve the flow of

communication in an organization and it allows

managers to discover the way the work is being done

in the informal way (Noel et al., 1979).

To re-engineer an enterprise organization,

(Boulmakoul and Besri, 2013) define a set of

operations applicable along two viewpoints:

Organizational and performer status. They show

how several framework and toolkit can be used for

process mining of the organizational perspective,

visualizing and analyzing the organizational

structure.

In (Hong et al., 2012), the authors present a

methodology to derive an organizational structure.

The methodology has four phases. The first one

collects source data from BPMN models measured

by transfer-of-work metrics; the metrics were

defined to derive relations between resources from

process logs (Van der Aalst et al., 2005) (Choi et al.,

2007), (Song and Van der Aalst, 2008).

In the second phase, the BPMN model is

transformed into a process network that is

diagnosed, in the third phase, by five problems-

oriented approaches: verticality of workflows,

degree of bottlenecks, core competence of business

processes, authority that corresponds to the position,

and degree of business cooperation.

The aforementioned works are based only on

knowledge rediscovery relying on the structural

aspect. They can identify central nodes in the

network and they can take measures over the

structure of the social network model such as node

centrality, node betwenness, density, geodesics

distance, diameter, connectivity of the graph, etc.

2.2 Discovery-based Approach

Besides the rediscovery approaches, other

approaches focused on discovering social network

knowledge through exploring the human perspective

of a group of models (Ahn et al., 2014).

More specifically, the authors in (Battsetseg et

al., 2013) (Kim, 2013) propose an approach for the

workflow-supported affiliation networking

knowledge discovery. They propose various

formalisms (Kim et al., 2014) and algorithms to

model, discover, and visualize the workflow

performer-role affiliation networking knowledge

from an Information Control Net (ICN) based

workflow model.

In the discovery-based approaches, the profile

information is typically represented as a matrix used

by algorithms to discover and analyse performer-

role affiliation networking and activity-performer

affiliation. In the affiliation network, performers are

linked through their joint participation in performing

roles. Conversely, roles are assigned to the

performers who are involved in the roles. Through

the performer-role affiliation networking knowledge,

it is possible to visualize in a workflow model how

performers and roles are simultaneous.

2.3 Discussion

Existing approaches deal with each type of

knowledge separately. However, using either social

network knowledge rediscovery or social network

knowledge discovery reduces the scope of the

information that can be extracted: An affiliation

presents a well-connected performer but not

necessarily similar in terms of theirs profiles. In

addition, an affiliation may be composed of similar

ICSOFT-EA2015-10thInternationalConferenceonSoftwareEngineeringandApplications

but loosely connected performers. In other words,

such separate use of the knowledge may lead to

inefficient restructuring solutions, which would

impact the organization performance.

3 STRUCTURE AND SEMANTICS

BASED CONTEXT

BPMN provides for the modelling of tasks assigned

to actors/performers. Hence, a BPMN business

process model

can be seen as a social network

model

()()

CEVG ,,

where:

−

()

EVG ,

is an undirected graph representing the

structure of the business process, with

being

the set of

nodes representing the performers

(i.e., the actors in the BPMN model), and

being the set of

edges connecting the nodes

(i.e., the flows in the BPMN model); and

−

is the semantic information of the network

model.

Table 1: Tabular representation of the semantic

information

in a social network model.

The semantic information

represents the

individual information of each performer in the

business process model: which tasks are performed

by the actor, and each actor's lane membership. It is

defined as a matrix of nodes x features. Each node

(i.e. performer) in the business process is described

by one row of features we call instance. The features

included in

may cover the functional,

informational, organizational and behavioural

contexts, or a combination of them. Table 1 shows a

tabular representation of an example of

The context concept is used to divide the set of

features into subsets according to different

perspectives

(Curtis et al., 1992). The features

provide for the discovery of unseen information

belonging to each perspective (functional,

informational, behavioural and organizational) and

related to each performer.

The functional perspective (Curtis et al., 1992)

represents what process elements are being

performed. The BPMN main concept that reflects

this perspective is Activity. In this perspective, the

feature that can be derived is IdTask, IdSubProcess.

In addition, since the informational perspective is

represented in terms of data (Curtis et al., 1992), the

data input and data output can be used as a set of

features. Furthermore, the organizational perspective

represents where and by whom process elements are

performed (Curtis et al., 1992). The main BPMN

concepts that reflect the organizational perspective is

Lane and Pool. The information that can be derived

from these BPMN concepts is IdLane and IdPool.

With the aforementioned concepts, we can now

define, for each node, the context which is a

particular set of values for each feature.

Definition 1: (Context

). Given a set of features

, a context

is one of the m-combinations of

the

elements of

. Note that

FC ∈

Definition 2: (Augmented social network model

): Given a social network model

()()

CEVG ,,

where

is the graph representing the structural

aspect,

the semantic aspect as a context, the

augmented social network model is defined as

()

ACGP

where

is the affiliation variable that

is derived from

and

The affiliation variable

of an augmented

social network

can be derived either from the

structural aspect, contained in

, or from the

semantic variables contained in

In the first

case, we assume that

; this means that the

determination of the affiliation

becomes a

general problem of graph clustering. Note that graph

clustering approaches use only the structure to find

cohesive groups. For

()

,VG

, only

is available,

the affiliation variable

can be generated using

traditional data clustering methods that use

(typically) vector representations of the data. Using

this data these methods produce groups of close

elements according some distance measure.

There is a gap between the available clustering

approaches designed for each one of these cases.

This gap opens a new study field, looking for new

ways to generate affiliation variables that integrate

the structural and the semantic aspect.

SemanticandStructuralPerformerClusteringinBPMNModelsTransformedintoSocialNetworkModels

4 CLUSTERING STRUCTURAL

AND SEMANTIC ASPECTS

The main objective of our work is to use both

structural and semantic aspects of a business process

model to restructure it based on the social network.

To do so, we need to generate the performer

partitions which are the result of a clustering

process. The obtained performer affiliation should

represent groups of well-connected and similar

performers.

Figure 1 presents the general diagram of our adopted

structural and semantics clustering.

Figure 1: Performer affiliation using the structural and

semantic aspects.

First, the context

from the social network

model is used to find an auxiliary affiliation of

performers based on the semantic aspect

SEM

. This

affiliation contains groups of similar performers. It

is obtained by the semantic information represented

by a proposed distance called Task-Lane

LPA

−−

This distance is used to determine the number of

tasks and lanes that are different for any pairs of

actors. In addition, to account for

the structural

aspect, the semantic distance

LPA

−−

is multiplied

by the structural distance

which expresses the

proportion of the sequence flow connecting the

performers. The integration of structural and the

semantic aspects produces a new performer’s

affiliation

STRSEM

−

that contains information from

both aspects.

Finally, in order to cluster the performers, we

adapt the hierarchical algorithm (Kantardzic, 2002)

to our domain to generate the partition groups of the

performers according to the similarities of their

features and the relationships between them. The

steps of our agglomerative hierarchical clustering

are presented in Algorithm 1. This algorithm has two

main advantages: it requires no a priori information

about the number of clusters required, and it is easy

to implement.

Algorithm 1

Let

{}

xxxxX ,....,,

321

be the set of data points.

1. Begin with the disjoint clustering having level

()

00 =L

and sequence number

0=m

2. Find the least distance pair of clusters in the

current clustering, say pair (r), (s), according

to:

()()

[]

()( )

[]

jidsrd ,min, =

where the minimum is over all pairs of clusters in

the current clustering.

3. Increment the sequence number:

1+= mm

Merge clusters (r) and (s) into a single cluster

to form the next clustering m. Set the level of

this clustering to

() ()()

[]

srdmL ,=

4. Update the distance matrix, D, by deleting the

rows and columns corresponding to clusters

(r)and (s) and adding a row and column

corresponding to the newly formed cluster.

The distance between the new cluster, denoted

(r,s) and old cluster(k) is defined in this way:

()( )

[]

()()

[]

()()

[]

skdrkdavergemeansrkd ,,,,, =

5. If all the data points are in one cluster then

stop, else repeat from step 2.

4.1 Example

To illustrate our approach, we will use the “supply

management business process” shown in Figure 2.

The task assignment to actors and lanes are listed

in Table 2. Table 3 shows the corresponding binary

affiliation matrix “Activity-Performer-Lane”.

Each row in Table 3 expresses a vector of

features representing one actor. In our example, we

aim to cluster the actors which belong to the same

pool: “supply management process’’. We used the

following features: IdTask to represent which tasks

are performed by the actor, and Id-Lane to express

each actor's lane membership.

An “Activity-Performer-Lane” (A-P-L)

affiliation network model is graphically represented

by a bipartite graph, and it is mathematically

represented by an affiliation matrix (see Table 3).

ICSOFT-EA2015-10thInternationalConferenceonSoftwareEngineeringandApplications

Figure 2: BPMN example: Supply management business process.

Table 2: Tasks assignment to actors and lanes.

Each entry x

i,j

of the A-P-L matrix is filled

according to the following rule:







otherwise

LlaneorTactivitiywithaffiliatedisAperformerif

jji

,,,1

Based on Table 3, we calculate the first distance D

A-

P-L

as the Euclidean distance between two actors

(two vectors in Table 3). D

A-P-L

determines the

number of tasks and lanes which are different

between a pair of actors. Table 4 shows the values

of the activity-performer-lane distance.

The second calculated distance is the flow

distance D

−

FTF

(1)

where: N

is the total number of sequence flows sent

directly from one actor to another, and N

is the

total number of sequence flows in the model. This

distance represents the distance between actors in

terms of how work is moved among them. The 1

added in the denominators is to avoid a division by

0. Table 5 lists the D

values for the running

example. Based on D

A-P-L

and D

, we calculate the

total distance as follows:

()

TFLPAF

ddDDD ++=

−−

(2)

We add

()

dd +

in formula (2), in order to avoid

the case of a null distance when D

=0 and

A-P-L

≠0

and conversely. We suppose that

=0.

SemanticandStructuralPerformerClusteringinBPMNModelsTransformedintoSocialNetworkModels

Table 3: Binary affiliation matrix “Activity-Performer-Lane” of Figure 2.

Table 4: Euclidian distance D

A-P-L

between actors.

Table 5: The D

distance between actors.

Table 6: Total distance D between actors.

Table 6 summarizes the total distance

for the

running example. This distance matrix is the input to

the

hierarchical clustering algorithm to determine

the actor affiliation.

To illustrate the application of this task, we next

show how the classification objective, making

homogeneous and distinct groups, can be

mathematically formalized by using the concepts of

intra-class inertia (Kantardzic, 2002). The goal is to

find the partition K classes whose inertia intra class

is minimal.

The inertia is defined as follows:

Let

is a group of individuals partitioned into nbg

classes

.....,

21 nbg

ggg

The intra-class inertia I is equal to:

()

!2!2

where

×−



























∈

nbg

gcc

ccd

nbg

(3)

4.2 Application of Hierarchical

Algorithm Clustering

Hierarchical clustering is a method of cluster

analysis that seeks to build a hierarchy of clusters.

We used the agglomerative strategy for hierarchical

clustering which a "bottom up" approach: each

observation starts in its own cluster, and pairs of

clusters are merged as one moves up the hierarchy.

The linkage criterion determines the distance

between sets of observations as a function of the

pairwise distances between observations. In our

example, we use the average linkage clustering. The

following steps are conducted over the running

example:

STEP1: Each observation is in its own cluster: {A1}

{A2}, {A3}, {A4}, {A5}, {A6}, {A7}. The input

distance matrix

(L = 0 for all the clusters) is the total

distance shown in Table 6.

In the first step, the inertia is equal to zero:

STEP2: Based on the input distance matrix, the

nearest pair of actors are (A5, A7), (A5, A6), and (A6,

A7). We select for example, A5 and A7, at distance

0.127. These actors are merged into a single cluster

called

"A5/A7". The level of the new cluster is L

(A5, A7) = 0.127 and the new sequence number is m

= 1.

Then we compute the distance from this new

compound object to all other objects. In average link

clustering the rule is that the distance from the

compound object to another object is equal to the

mean average distance from any member of the

cluster to the outside object. So the distance from

"A5/A7"

to A6 is chosen to be 0.127, which is the

average distance from A5 to A6, and A6 to A7.

ICSOFT-EA2015-10thInternationalConferenceonSoftwareEngineeringandApplications

After merging A5 with A7

we obtain the

following matrix representing the clusters:

{A1} {A2}, {A3}, {A4}, {A5, A7}, {A6}

Table 7: Distance matrix for step 2.

We calculate then the inertia that corresponds to

this step: I

=0.0026.

STEP3: In this step, because

min d(i,j)=d((A5/A7),A6) = 0.127, then we merge

"A5/A7"

and A6 into a new cluster called {A5,

A6,A7}, which gives us L((A5/A),A6)=0.127, m=2

distance matrix shown in Table 8.

Table 8: Distance matrix for Step 3.

The derived clusters are: {A1}, {A2}, {A3},

{A4}, {A5, A7, A6} and the inertia is I

=0.0032.

STEP4: Because we have

min d(i,j)=d(A1,A2) = 0.16, then we merge A1/A2

into a new cluster called {A1,A2}. At the end of this

step, we have L(A1,A2) = 0.16, m=3 and the

distance matrix shown in Table 9.

At the end of this step, the obtained clusters are:

{A1, A2}, {A3}, {A4}, {A5, A7, A6} and the inertia

is I

=0.009.

Table 9: Distance matrix for step 4.

STEP5: Because we have min d(i,j)=d((A1/A2),

A3)=0,251, then we merge

A1/A2 with A3 into a

new cluster called {A1,A2,A3}. Thus, we have:

L(A1/A2)=0.251, m=4

and we obtain the distance

matrix of Table 10.

Table 10: Distance matrix for step 5.

At this step, the obtained clusters are:{A1, A2, A3},

{A4}, {A5, A7, A6} and the inertia is: I

=0.019.

STEP6: min d(i,j)=d((A1/A2/A3),A4) = 0.257

which

leads to merging A1/A2/A3 with A4

into a new

cluster called {A1, A2, A3, A4}.

L((A1/A2/A3),A4) = 0.257, m=5

After merging A1/A2/A3 with A4 we obtain the

distance matrix of Table 11, the custers: {A1, A3,

A2, A4}, {A5, A7, A6} with an inertia I

=0.034.

Table 11: Input distance matrix for step 6.

STEP7: Finally, we merge the last two clusters at a

level of 5.2. As depicted in figure 3, the inertia

reaches its highest value in this step. We can see that

the difference between the inertia values in two

consecutively steps increases from step 5 to step 6.

The obtained result shows that the difference

between the inertia in a time (t) and (t-1) must not

exceed

015.0=

() ( )

≤−

−tt

Figure 3: The inertia curve during the six iterations.

In this example, the clustering at step 5 is

considered optimal: {A1, A2, A3}, {A4}, {A5, A7,

A6}. Based on this clustering, we obtain three lanes:

the first lane contains the actors {A1, A2, A3}. The

second lane contains the actor {A4} and the last lane

contains the actors {A5, A7, A6}.

SemanticandStructuralPerformerClusteringinBPMNModelsTransformedintoSocialNetworkModels

4.3 Experimental Evaluation

To evaluate the obtained inertia threshold, we

worked on forty business processes models. In this

empirical study, we applied the hierarchical

algorithm to forty business process models, and we

calculated the inertia for each case. The results

showed that the best clustering is obtained in 36

models with a threshold inertia value that does not

exceed 0.015.

5 CONCLUSIONS

The information contained in a socio-semantic

network is tied both to certain features pertinent to

individual performers (semantic information) and

their organizational relationships (structural

information). Such information allows to perform

more comprehensive analyses over the network from

different perspectives, which provides for better

restructuring decisions.

Unlike existing the approaches which use one

type of information, in this paper, we proposed an

approach for social network restructuring that uses

both structural and semantic information. Our

approach relies on the definition of the concept of

context which augments the social network with

semantics pertinent to the business process. In

addition, it uses two new distances that account for

the semantic and structural information, and applies

a hierarchical clustering algorithm to identify

performer clusters. Each cluster represents an

affiliation of well-connected performers that have

similar profiles.

We are currently defining a graph-based method

that uses the obtained clusters to restructure an

organization. This method extends our preliminary

identified set of rules for transforming a BPMN

model into a behaviourally equivalent one (Khlif and

Ben-Abdallah, 2015).

REFERENCES

Ahn, H., Park, Ch., Kim, K.P., 2014. A correspondence

analysis framwork for workflow-supported performer-

activity affiliation networks. In ICACT'14, 16th Intern.

Conf. on Advanced Communication Technology. pp.

350-354. IEEE.

Battsetseg, A., Hyun, A., Lee, Y., Park, M., Kim, H.,

Yoon, W., Kim, K. P., 2013. Organizational Closeness

Centrality Analysis on Workflow supported Activity-

Performer Affiliation Networks. In ICACT’13, 15th

International Conference on Advanced

Communication Technology. pp. 154-157.

Boulmakoul, A., Besri, Z., 2013. Scoping Enterprise

Organizational Structure through Topology

Foundation and Social Network Analysis. In

INTIS’13, Innovation and News Trends in Information

Systems. Tanger, Morocco.

Choi, I., Song, M., Kim, K.P., and Lee, Y., 2007. Analysis

of social relations among organizational units derived

from process models and redesign of organization

structure. In Journal of the Korean Institute of

Industrial Engineers. 33(1) pp.11-25.

Curtis, B., Kellner, M., Over. J., 1992. Process Modeling.

Communication of the ACM. 35(9).

Hong, S., Lee, Y., Kim, J., Choi, I., 2012. A Methodology

for Redesigning an Organizational structure based on

Business Process Models using SNA Techniques. In

Journal of Innovative Computing, Information and

Control. 8(B), ISSN 1349-4198. pp. 5411-5424.

ISO/IEC 19510, 2013. Information technology -- Object

Management Group Business Process Model and

Notation. Available from: http://www.iso.org/iso/

catalogue_detail.htm%3Fcsnumber%3D62652.

Kantardzic, M., 2002. Data Mining: Concepts, Models,

Methods, and Algorithms. Wiley-IEEE Press.

Khlif, W., Ben-Abdallah, H., 2015. Integrating semantics

and structural information for BPMN model

refactoring. In SERA'15, 13

Intern. Conf. on Software

Engineering Research, Management and Applications.

Toulouse.

Kim, H., Ahn, H., Kim, K.P., 2014. Modeling,

Discovering, and Visualizing Workflow Performer-

Role Affiliation Networking Knowledge. In TIIS’14,

Journal of Transactions on Internet and Information

Systems. 8(2). pp. 689-706.

Kim, K.P., 2013. Discovering Activity-Performer

Affiliation Knowledge on ICN-based Workflow

Models. In Journal of information science and

engineering. 29. pp. 79-97.

Noel, M. T., Micheal, L. T., Charles, F., 1979. Social

Network Analysis for Organizations. In The Academy

of Management Review. 4(4), pp. 507-519.

Oinas-Kukkonen, H., Lyytinen, K., Yoo, Y., 2010. Social

Networks and Information Systems: Ongoing and

Future Research Streams. In Journal of the

Association of Information Systems. 11(2). pp. 61-68.

Stanley, W., Katherine, F., 1999. Social Network

Analysis: Methods and Applications. In Revue

française de sociologie. 36-4. pp. 781-783.

Song, M., Van der Aalst, W. M. P., 2008. Towards

comprehensive support for organizational mining. In

Decision Support Systems. 46(1), pp.300-317.

Van der Aalst, W. M. P., Reijers, H. A., Song, M., 2005.

Discovering social networks from event logs. In

Computer Supported Cooperative Work. 14(6), pp.

549-593.

ICSOFT-EA2015-10thInternationalConferenceonSoftwareEngineeringandApplications