EVARES: A Quality-driven Refactoring Method for Business Process

Models

Wiem Khlif

, Nouchène Elleuch Ben Ayed

and Hanêne Ben-Abdallah

2,1

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

King Abdulaziz University, K.S.A.

Keywords: BPMN Models, Transformation Rules Ordering, Perspectives, Quality Metrics, Quality Sub Characteristics.

Abstract: The business performance of an enterprise tightly depends on the quality of its business process model (BPM).

This dependence prompted several propositions to improve quality sub-characteristics (e.g. modifiability and

reusability) of a BPM through transformation operations to change the internal structure of the model while

preserving its external behaviour. Each transformation may improve certain metrics related to one quality sub

characteristic while degrading others. Consequently, one challenge of this model transformation-based quality

improvement approach is how to identify the application order of the transformations to derive the “best”

quality model. This paper proposes a local optimization-based, heuristic method to decide on the application

order of the transformations to produce the best quality BPM. The method is guided by both the perspectives,

and the impact of each transformation on the quality metrics pertinent to the perspectives as well as the quality

sub characteristics of interest to the designer. The method’s and an experimental evaluation are presented.

1 INTRODUCTION

To improve the performance of its business process,

an enterprise often needs to restructure its Business

Process Model (BPM). To provide for model

restructuring, several refactoring techniques have

been proposed, cf. (La Rosa et al., 2011). These

techniques are transformation-based and structural

pattern-driven, and they restructure a model without

changing its external behaviour. In addition, they are

quality focussed to assist business analysts to improve

quality sub-characteristics of the BPM like

understandability, reusability, and modifiability. For

example, several works (e.g. (La Rosa et al., 2011))

rely on the empirically shown fact that their

transformations can lead to “better structured”

models.

The model transformation-based approach to

improve the quality of a BPM faces two main

challenges: completeness of the transformation

operations, and identification of their application

order which produces the best quality model. The

second challenge is the focus of this paper where the

quality sub characteristics are assessed through a set

of BPM metrics.

The final quality of a BPM depends on the order

of application of the transformations for two reasons:

On the one hand, a transformation may have

conflicting impact on quality metrics and thus sub-

characteristics; on the other hand, being structural

pattern-based, the application of a transformation

enable and/or disable other transformations.

Evidently, with a large set of transformations, it is

impractical to try all possible (exponential number of)

combinations of transformations to identify the “best”

quality model. Evidently, an ad hoc application

approach defeats the restructuring purpose.

Face to this challenge, the literature is rather

silent. In fact, this challenge is addressed only by

(Fernández-Ropero et al., 2013) who statistically

proposes to apply the transformation categories in a

particular order; but, within one category, the

transformations are still applied in an ad hoc way. In

addition, none of the existing transformation-based

works, e.g. (Fernández Ropero et al., 2013), considers

the gain of transformation-based refactoring

techniques in terms of business perspectives and/or

quality sub-characteristics.

This paper proposes a new approach to tackle this

challenge within the method EVARES (EVAluation

and REStructuration of BPMN models) (Khlif et al.,

2017). EVARES is a quality-driven and

transformation-based method to refactor BPMN

Khlif, W., Ayed, N. and Ben-Abdallah, H.

EVARES: A Quality-driven Refactoring Method for Business Process Models.

DOI: 10.5220/0006315504090416

In Proceedings of the 19th International Conference on Enterprise Information Systems (ICEIS 2017) - Volume 3, pages 409-416

ISBN: 978-989-758-249-3

409

models. Its refactoring operations were determined

based on a set of structural patterns that we identified

empirically. It assesses quality in terms of a set of

pertinent metrics (e.g., CW, CFC, TNG, NSF, Den,

NOA, etc (Cardoso et al., 2006).

This paper enhances EVARES with a local

optimization-based, heuristic algorithm to decide on

the application order of the transformations to

produce the best quality BPM (Section 2). The

algorithm is guided by the perspectives (functional,

organizational, informational, behavioural), and the

impact of each transformation on the quality metrics

pertinent to the perspectives as well as the quality

sub-characteristics of interest to the designer. To do

so, we identify, for each transformation, the set of

modelling metrics that it affects. In addition, we

classify the transformations according to business

process perspectives and quality sub characteristics

(modifiability, comprehension and reusability)

(ISO/IEC25010, 2011). Besides presenting the

algorithm, this paper also presents how EVARES

assesses the quality of a BPM (Section 3) and the

EVARES tool (Section 4). Finally, Section 5 presents

related works and outlines future work.

2 RULE APPLICATION ORDER

IDENTIFICATION

EVARES (Khlif et al., 2017) is a method for

restructuring BPMN models based on semantic and

structural information. It operates in two phases:

restructuring followed by evaluation (Section 4).

The EVARES restructuring phase is driven by 28

transformation rules which we identified based on a

set of structural patterns we determined empirically

(Khlif et al. 2017). To facilitate their application, the

transformation rules operate on canonical fragments

that can be determined by the algorithm proposed in

(Polyvyanny, 2012) to decompose a BPMN model

into two special kinds of process fragments: Single

Entry Multiple Exit (SEME) to apply the behavioral

and informational rules, and Single Entry Single Exit

(SESE) fragments to apply the organizational and

multi-perspective rules. The selection of the

transformation rules is driven by the designer’s

perspective(s) of interest. Thus, we classify EVARES

transformation rules into organizational, functional,

behavioral, informational and multi-perspectives.

Due to space limitation, we present six rules (Khlif et

al., 2017) that we will illustrate, in section 4, through

the ‘Loan process’ example model.

R1-beh: If an exclusive gateway has fan-outs to two

parallel gateways G1 and G2 which are linked

respectively to activities A,B and A, C, then link B

and C to the exclusive gateway which will be linked

to A by a parallel gateway.

R2-Org: Merge directly connected activities

performed by two actors in the same lane and

associate the resulting activity with the actor who has

permission to perform the original activities.

R3_Org: Duplicate an activity in two lanes if it is

followed by a parallel fragment that is performed by

actors in the two lanes, and these actors have the

permission to perform the first activity.

R4-Org: If a lane contains only an activity

respectively followed or it is between two parallel

fragments which are performed by actors in different

lanes and who have the permission to perform the first

activity, then apply successively the following rules:

R3-Org, R2-Org.

R3_Org and R4_Org can also be applied in one lane.

We call, respectively, these variants R3_Org_V and

R4_Org_V. In this case, R4_Org_V applies

successively R3_Org_V and R2_Org.

R5_Inf: If there is more than one end event in

different lanes, then all end events will be grouped

with an exclusive, inclusive or parallel gateway,

depending on the initial structural context.

R6_Multi: If an inclusive fragment is attached to two

exclusive fragments containing a duplicated task,

then associate it to the actor who has the permission

to perform it.

Table 1 summarizes the effects of these rules on

the metrics. The minus sign (-) means the metric

should be minimized to improve the model, while the

plus sign (+) means the metric should be maximized

to improve the model quality; the sign NA means that

the metric is not affected by the rule.

We propose a greedy algorithm for the rule

application order identification problem (see

Algorithm 1). The algorithm expresses an optimum

local choice in the hope to produce a global

optimization. Once made, a choice cannot be

unperformed, even if, in one step, this choice is

detrimental to the production of an optimal solution.

ICEIS 2017 - 19th International Conference on Enterprise Information Systems

410

Table 1: Trends of quality metrics when applying the transformation rules.

Perspectives

Rules

Complexity and coupling metrics

CFC

NSF

TNG

NOA

NSFG

TNE

Den

AGD

MGD

Behavioural

R1-Beh

Organizational

and functional

R2-org

R3-org

R3_Org_V

R4-org

R4_org_V

Informational

R5_Inf

Multi-

perspectives

R6_Multi

Algorithm 1

Input: Model M, a Set of Rules SR, Choice

selected perspectives, a set of Metrics

SMeas divided to SMeas

to minimize and

SMeas

to maximize

Output: Metric Values newVMeas, a Set

of Selected Rules SRF to apply, and the

restructured model M1

1. Main(){

2. GQMcalGlobalQuality(M);

3. F[]decompose(M);

4. K1

5. While ( not empty(F[])){

6. Max0; Min99999;

7. For (j=1; j<=count(F[]); ++j) {

8. SRF[k][j]identifyRules(Choice[],SR[],F

[j]);// SRF: a set of rules to apply

based on the chosen perspective(s)

9. ARF[k][j]chooseApplicableRule

(SRF[k][j], F[j], M);// ARF: set of

applicable rules

10. M1transform(M, ARF[k][j]);

11. GQM[j]calGlobalQuality(M1);

12. If SMeas==SMeas

then {

13. if (max<GQM[j]) then{ maxGQM[j];

14. BRF[k] ARF[k][j];}} //this is:

best rule to apply in this iteration

15. else if (min>GQM[j]) then{minGQM[j];

16. BRF[k] ARF[k][j];}} }

17. M transform(M, BRF[k]);

18. F[]decompose(M); K++;}}

19. Function chooseApplicableRule(ARF,

F, M) return string{

20. VMeas[]calMeasuresValues

(SMeas, M);

21. If (ARF is empty) then exit;

22. Else For each rule R in ARF do

23. {Mtransform(M,R);

24. newVMeas[]=calMeasuresValues

(SMeas, M);

25. If SMeas==SMeas

then{ //R invokes

measures to minimize

26. flagTrue;

27. For (i=1; i<=count(VMeas);i++){

28. If(NewVmeas[i]>Vmeas[i])then

29. flagfalse; end if; }

30. If (flag==True) then return R;

31. Else{ // Compare gains to losses

32. gaincomputeGain(VMeas[],newVMeas[],

"to_minimize");

33. LosscomputeLoss(VMeas[],newVMeas[],

"to_minimize");} End if;

34. If gain>loss then return R;

35. Else if SMeas==SMeas

then{// R

invokes only measures to maximize

36. flagTrue;

37. For (i=1; i<=count(VMeas);i++){

38. If (NewVmeas[i]<Vmeas[i])then

39. flagfalse; end if;}

40. If (flag==False) then // Compare

gains to losses

41. {gaincomputeGain(VMeas[],newVMeas[]

,"to_maximize");

42. LosscomputeLoss(VMeas[],newVMeas[]

,"to_maximize");}End if;

43. If gain>loss then return R;}

44. Else // R invokes mixed measures

45. {NSum0; Sum0;

46. For (i=1; i<=count(VMeas);i++){

47. NSum+=NewVmeas[i]; Sum+=Vmeas[i];}

48. If NSum/Sum<=1 then return R;}}

49. Function calGlobalQuality(M) {

50. Var total=0;

51. VM[]calMeasuresValues(SMeas, M);

52. For(i=1;i<=n;++i){

53. total=+VM[i]*VM[i];}

54. return (square(Total));}

55. Function computeGain(VMeas[],

newVMeas[,objective)return number{

56. NSum0; Sum0;

57. For (i=1; i<=count(VMeas);i++){

58. NSum+=NewVmeas[i];Sum+=Vmeas[i];}

59. If (objective=="to_minimize") then{

60. Return (1-(NSum/Sum));}

61. Elseif (objective=="to_maximize")

then { return (1-(Sum/NSum));}}

62. Function computeLoss(VMeas[],

newVMeas[],objective) return number{

63. NSum0; Sum0;

64. For (i=1; i<=count(VMeas);i++){

65. NSum+=NewVmeas[i];Sum+=Vmeas[i];}

66. If (objective=="to_minimize") then{

67. Return (1-(Sum/NSum));}

68. Else if (objective=="to_maximize")

69. then {Return (1-(NSum/Sum));}}

EVARES: A Quality-driven Refactoring Method for Business Process Models

411

As described in Algorithm 1, the heuristic

identifies, for each instance M

of the model in an

iteration i, all rules applicable to the fragments

identified in M

while respecting the perspectives

chosen by the designer (lines 5-8). Then, the

algorithm examines, for each rule, if it improves the

overall quality (See Exp. 1) of a model M

in order to

determine their applicability and retain it as candidate

(lines 19-48). This decision takes into account how

the metrics affected by each rule will be interpreted to

obtain a good quality model. This leads to three

possible cases: all metrics must be minimized (lines

25-34), all metrics must be maximized (lines 35-43),

or the metrics are the mixture of the two cases (line

44-48). In each case, the rule is retained only if the

new metrics’ values are not against their tendency as

shown in Table 1.

If the designer chooses a perspective and there is

no applicable rule, the algorithm considers all rules

on an equitable basis. Once the rules to be applied are

identified, the algorithm selects the best rule that

improves the overall quality lines (12-17).

The overall quality is calculated based on all

metrics that are assumed to be minimized or

maximized. Given the metrics and their values,

,...v

, the overall quality of a model M is defined as

follows:

 

v+....+ v+v

(1)

Note that Algorithm 1 stops if the model M

does not

contain any fragments or no rule is applicable.

3 EVARES QUALITY

EVALUATION

EVARES proposes an evaluation method based on a

quantitative measurement and an interpretation. The

evaluation method can be driven by either the

perspectives or the quality sub-characteristics.

3.1 Perspective-driven Measurement

and Interpretation

In EVARES, the measurement activity pilots the

restructuring process based on the selected

perspectives, and it estimates the added value of the

restructuring process. The quality metrics

measurement compares the initial BPMN model and

the restructured one. It calculates all metrics and

classifies them in the two levels (Khlif et al., 2017).

At the first level, the metrics belong to one of three

categories: complexity, coupling and cohesion. At the

second level, the metrics of each category are

associated to perspectives.

Each time a rule relative to the designer’s

perspective is identified as applicable, the values of

all its impacted metrics are calculated and saved in a

report classified according to its quality sub-

characteristics. The interpretation uses the initial

model report (M

) and the restructured one (MR). It

calculates the improvement ratio RI

(lines 55 -69).

A positive RI

indicates that the transformation

improves the model quality; otherwise, it degrades it.

3.2 Quality Sub-characteristics-driven

Measurement and Interpretation

Even when the calculated improvement Ratio (RI)

indicates an improvement in the restructured model

based on the business/organizational goals, what

about the model’s intrinsic quality? That is, what is

the impact of restructuration based on quality sub-

characteristics such as understandability,

modifiability and reusability? The model quality

depends on the metrics’ tendency associated with

each sub-characteristic.

3.2.1 Case of Metrics to Minimize or

Maximize

The measurement activity produces for each quality

sub-characteristic the associated metrics and their

values. The interpretation of the quality metrics gives

an evaluation of the business process quality based on

the following information for each metric:

 a priority order between the metrics based on

their use frequency. It is deduced from the

literature (Cardoso et al., 2006).

 a threshold interval that reflects the optimal

value of a metric. It is the result of empirical

studies (Sánchez-González et al., 2010).

Given the priorities and the values, the

interpretation activity determines the sum



weighted by the priorities and the metrics values

associated to the same quality sub characteristic

Formally, for each sub-characteristic

, the global

quality model



is calculated as follows:

 Let

,..

the priorities assigned to each

metric in

.The weight of each metric is:







(2)

ICEIS 2017 - 19th International Conference on Enterprise Information Systems

412

 Let

,..

be the metrics’ values determi-

ned by the measurement activity. Then:

k ik ik







(3)

After that, we calculate the sum

mink



(resp.

maxk



) of

the metrics threshold minimal values (resp. maximal

values) in a sub-characteristic

. These values are

weighted by the metrics priorities



mink



maxk



are calculated as follows:

 Let

minik

and

maxik

the minimal and maximal

thresholds associated to the metrics. Then:

min min

k ik i k







(4)

max max

k ik i k







(5)

The comparison between

mink



maxk



provides an

assessment (efficient, medium or inefficient) of the

model according to the sub-characteristics

min maxk k k

  



(6)

3.2.2 Case of Mixed Metrics

To assess the BPM quality in the case of mixed

metrics (i.e., they have tendencies to be minimized

and maximized) in the same sub-characteristic, we

use fuzzy logic (Zadeh, 1965) which is based on

"degrees of truth". Fuzzy sets have elements with

degrees of membership. In fact, we calculate, for each

category of metrics’ tendency, the membership

degrees of the metrics.

 For the metrics that should be maximized :

We calculate the membership degree

associated

to each metric m

that should be maximized:

min

max min







(7)

Where

: the metric value,

min

: the metric threshold minimal value

max

: the metric threshold maximal value

Fuzzy logic allows different degrees of response to

the question "Is the metric value high? "(Figure 1) :

Figure 1: Membership function of a fuzzy set of metrics that

should be maximized.

 Below or equal

min

, the metric value is

high with a 0% confidence level.

 Above or equal to

max

, the metric value is

high with a 100% confidence level.

 In

 

min max

,vv

, the metric value

has a

high value with a

confidence.

 For the metrics that should be minimized:

We calculate the membership degree

associated

to each measure

that should be minimized:

max

max min







(8)

Where

: metric value,

min

: metric threshold minimal value

max

: metric threshold maximal value

Figure 2: Membership function of a fuzzy set of metrics that

should be minimized.

Fuzzy logic also allows to respond to the question

"Is the metric value minimal? "(Figure 2) :

 The metric value is minimal with 100%

confidence level if it is equal or below

min

 The metric value is minimal with 0%

confidence level if it is equal or above

max

 In

 

min max

,vv

, the metric value

is minimal

with a

confidence degree.

In each case (metrics that should be maximized or

minimized), the membership degree is equal to:

EVARES: A Quality-driven Refactoring Method for Business Process Models

413

Figure 3: BPMN example: Loan process.

q 

then

q 

(9)

q 

then

q 

(10)

Let

,..

k nk

the membership degrees of metrics that

belong to a sub-characteristic, then we calculate the

average values

obtained from metrics values that

should be minimized and maximized:





(11)

We assess the quality model as follows:

 If

q 

then the model is very efficient;

 If

q 

then the model is inefficient;

 If

 

0,0,5

q 

then the model is medium;

 If

> 0,5 and

<1 then the model is

efficient

To facilitate the application and evaluation of our

method, we have implemented a tool named

EVARES Quality as an Eclipse TM plug-in (Eclipse,

2011). It is composed essentially of two components:

1) The BPMN model restructurer contains the Rule

application order identification and Transformer; and

2) The BPMN quality evaluator contains the

Calculator and the Interpreter. In order to illustrate the

functioning of our tool, we apply it to the “Loan

process” model shown in Figure 3.

4 EVARES QUALITY TOOL

4.1 BPMN Model Restructuring

Once the designer chooses his/her perspective(s) of

interest, the rule application order identification

module displays the set of rules corresponding to all

combinations of the chosen perspectives (Figure 4).

For instance, in the case of organizational and

informational perspectives, the heuristic promotes at

each iteration, the applicable organizational and

informational transformation rules that will be

retained even if they don’t produce the best model

quality. By selecting the "Show details" button in

Figure 4, the rule application order identification

represented by Figure 5 will display the result of the

running example. In this GUI, the designer retains

those rules he/she thinks are convenient. In the

running example, suppose we retain the informational

Figure 4: All possible combinations based on the designer’s

objectives.

ICEIS 2017 - 19th International Conference on Enterprise Information Systems

414

Figure 5: Rules order applied to "Loan Process" for the

informational and organizational perspectives.

rule R5_Inf even though it increases the quality

model. Based on the selected rules, the tool displays

the applicability order of rules that provides the best

quality by favouring the chosen perspective(s). In the

example, the best order is: R2_Og, R4_Og, R5_Inf,

R_Lit_XOR, R1_Beh. Figure 6 presents the model

after transformation.

4.2 BPMN Quality Evaluator

The metrics calculator produces the metrics values

before and after the transformation of the model.

Afterward, the interpreter calculates the ratio

improvement (RI) of metrics, and compares the

obtained results to threshold values of metrics

deduced from empirical studies (Sánchez-González

et al., 2010). Some threshold are introduced by the

user with stars (*). Based on the quality sub-

characteristic, the interpreter gives an evaluation of

the business process under analysis (Figure 7).

Figure 7: Quality model evaluation.

Figure 6: Loan Process after the last transformation.

EVARES: A Quality-driven Refactoring Method for Business Process Models

415

5 RELATED WORKS AND

CONCLUSION

Refactoring/transformation-based approaches to

improve the quality of BPM stand on three pillars:

quality assessment means, refactoring operations, and

their application strategy.

For model quality assessment, our method

EVARES relies on a set of metrics mapped to quality

sub-characteristics (ISO/IEC 25010, 2011). It assess

more quality sub-characteristics than existing

propositions, e.g., (Fernández-Ropero et al., 2013)

cover understandability and modifiability whereas

Rolon et al. (Rolon et al., 2015) evaluate usability and

maintainability. In particular, this paper showed how

EVARES uses metrics to assess understandability,

modifiability and reusability. In addition, EVARES

characterizes the metrics’ tendency for each quality

sub-characteristic.

As for the second pillar, several researchers

proposed refactoring operations (La Rosa et al.,

2011), e.g., R-lit-XOR that replaces two or more

nested gateways of the same type with a single one.

EVARES offers transformations that account for

both the structural and semantic information, which

more open quality improvement opportunities. In

addition, EVARES classifies the proposed

transformations into the perspective(s).

Finally, except for (Fernández-Ropero et al.,

2013), none of the proposed works define an

application order strategy for their transformations.

Indeed, the authors use a statistical approach to

identify the best order of applying three categories of

refactoring operators (i.e., irrelevant, granularity and

completeness). To do so, they first propose six

execution orders of operators. Second, they execute

the six orders and collect the metrics’ values for each

BPM. Finally, they apply a univariant general linear

model test on the collected values to show that one

particular order best improves understandability and

modifiability: reducing the granularity, then

removing irrelevant elements. Nonetheless, in each

category, the transformation order is left undefined.

This statistical approach of identifying the

transformations’ application order is impractical for a

large number of transformations.

We by passed these difficulties by adopting a

heuristic approach that accounts for the metrics’

tendency. More specifically, we presented a heuristic,

greedy algorithm that, iteratively, selects applicable

transformations in order to optimize locally the model

according to both the designer’s perspectives and

quality sub-characteristics.

Evidently, our heuristic approach to the identification

of transformation application order operates through

a local optimization technique whose result depends

tightly on the correlation among the rules. Hence, our

future work focuses on analyzing the correlations

among the transformation rules. In addition, we will

examine restructuring BPM that is based on temporal

constraints.

REFERENCES

Cardoso, J., Mendling, J., Neuman, J., Reijers, H.A., 2006.

A discourse on complexity of process models. In

BPM’06, Intern. Workshop on Business Process

Mamagement. LNCS, Berlin, Springer, pp. 115-126.

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

Communication of the ACM, 35(9).

Dijkman, R., Gfellerb, B., Kusterb, J., Volzerb, H., 2011.

Identifying refactoring opportunities in process model

repositories. In Journal of Information and Software

Technology, 53(9), pp.937-948.

Eclipse Specification. (2013). Available from:

http://www.eclipse.org/.

Fernández-Ropero, M., Pérez Castillo, R., Cruz-Lemus,

J.A., Piattini, M., 2013. Assessing the Best-Order for

Business Process Model Refactoring. In (SAC’13),

proc. of 28th Symposium On Applied Computing.

Coimbra. Portugal, pp. 1400–1406.

ISO/IEC 25010. 2011. Software and system engineering–

Software product Quality Requirements and Evaluation

(SQuaRE)– Guide to SQuaRE. .

ISO/IEC 19510. 2013. Object Management Group BPM

http://www.iso.org/iso/catalogue_detail.htm%3Fcsnu

mber%3D62652.

Khlif, W., Ben-Abdallah, H., Elleuch, B. N, 2017. A

methodology for the semantic and structural

restructuring of BPMN models. BPM’17, Journal of

Business process management, Vol 23 (1), pp.16 – 46.

La Rosa, M., Wohed, P., Mendling, J., ter Hofstede,

A.H.M., Reijers, H. A., Van der Aalst, W.M.P., 2011.

Managing process model complexity via abstract

syntax modifications. In IEEE Transactions on

Industrial Informatics. 7(4), pp. 614–629.

Polyvyanny, A. 2012. Structuring process models.

Dissertation, University of Potsdam, Potsdam,

Germany, January.

Rolón, E., Chavira, G., Orozco, J., Soto, J.P., et al. 2015.

Towards a framework for evaluating usability of

business process models with BPM in health sector.

AHFE’15, 6

Conf. on Applied Human Factors and

Ergonomics. Vol 3, pp. 5603-5610.

Sánchez-González, L., García,F.,Mendling,J., Ruiz, F.

2010. Quality Assessment of Business Process Models

Based on Thresholds. In CoopIS’10, 18

Conf. on

Cooperative Information Systems, pp. 78-95.

Zadeh, L.A., (1965). Fuzzy sets. In Information and

Control. 8 (3): 338–353.

ICEIS 2017 - 19th International Conference on Enterprise Information Systems

416