A Step Forward in Semi-automatic Metamodel Matching: Algorithms and Tool

D. LopesA Step Forward in Semi-automatic Metamodel Matching

A Step Forward in Semi-automatic Metamodel Matching: Algorithms and Tool

A Step Forward in Semi-automatic Metamodel Matching: Algorithms and Tool

1

José de Sousa Jr1, Denivaldo Lopes1, Daniela Barreiro Claro2 and Zair Abdelouahab1

1 Federal University of Maranhão - UFMA, São Luís - MA, Brazil2 Federal University of Bahia - UFBA, Salvador - BA, Brazil

UFMA UFBA

ICEIS 2009


Agenda

• Introduction– Context

– Problem

– Motivation

• An Approach for Metamodel Matching– Foundation for Metamodel Matching

– Another Algorithm for Metamodel Matching

2


• Extending and Adapting SAMT4MDE– Modeling

– Prototyping

– Tests

• Conclusions


Introduction

• Context– The complexity of producing software systems has

increased due:

• Continuous evolution of the requirements.

• The creation of new technologies.

• Integration with legacy systems.

3

– The phases of software development, maintenance

and evolution become more difficult.

– Model Driven Engineering (MDE) has made the

management of this complexity possible.


Introduction

• Context– In MDE context, models are formal entities that can

be understood and processed by computers.

– Examples of MDE:

• Model Driven Architecture (MDA) from OMG.

4

• Eclipse Modeling Framework (EMF) from Eclipse Project.

• Software Factories from Microsoft.

– In model driven approaches, model transformation is

the basic operation to manage models.

Model A Model B

Transformation Definition

Transformation Engine

execute

input output

Main point to

achieve

transformation.


Introduction

• Context– However, the manual creation of transformation

definitions is a programming activity.

• It is error-prone and a tedious task.

– We need more than transformation definition.

5

– We need more than transformation definition.

The insight is to apply

MDE concepts to

develop MDE.


Introduction

• Context

6

* Lopes, D.: Study and Applications of the MDA Approach in Web Services Platforms, Ph.D.

Thesis, University of Nantes (2005)

Mapping Specification and Transformation Definition


Background

• Context

(or mapping specification)

contains the correspondences

between the source and the

Platform-Independent

Model (PIM)

7



between the source and the

target metamodel.

(or transformation definition)

contains the operational rules

to transform a model into

another.


Platform-Specific

Model (PSM)


Background

• Context

However, create manually

mapping models is not easy.

8



mapping models is not easy.

How support the semi-automatic

creation of mapping model (or

mapping specification)?

How match two metamodels?



Introduction

• Problem

– How support the semi-automatic creation of

mapping specification in Model Driven

Engineering (MDE)?

9

• Motivation

– Schema matching was applied with success

in other fields :

• Database integration and evolution.

• E-business.

• Data warehousing.


Agenda


– Problem

– Motivation



10



– Prototyping

– Tests

• Conclusions


An Approach for Metamodel Matching

• Foundation for Metamodel Matching

– A transformation definition can be defined as:

( )( ) bcMMa MsMMCMsMTransfba

/)(/,/ 21 →→

11

c

bacMM

b

a

M

MMMC

MsM

MsM

where

ba

metamodel theusing

created and between model mapping theis /

, metamodel theusing created system afor model a is

, metamodel theusing created system afor model a is

2

1

→




– The operator match can be defined as:

( )ba MMba CMMMatch

→=,

12

How implement this operator?

What are the features to be considered in

the algorithm executed by this operator?


≠−

≅

=

=

xi

xi

xi

xi

ba

ba

ba

ba

11

11

11

11

if 1

if 0

if 1

),(ϕ

= ba if 1

1. Create C.

2. Initialize C.

3. Find leaf classes.

4. Select equal or similar classes

5. After iterating (4) to a fix point, put each c1p in C.

ϕ is calculated using dictionaries,

taxonomies, crosskind



– An initial algorithm for metamodel matching*.

OK!

This work but is not enough!

We need more than metamodel

13

≠−

≅

=

=

yj

yj

yj

yj

ba

ba

ba

ba

22

22

22

22

if 1

if 0

if 1

),('ϕ

≠−

≅

=

=

zk

zk

zk

zk

ba

ba

ba

ba

33

33

33

33

if 1

if 0

if 1

),(''ϕ

5. After iterating (4) to a fix point, put each c1p in C.

6. Select equal or similar data types

7. After iterating (6) to a fix point, put each c2q into C.

8. Select equal or similar enumerations

9. After iterating (8) to a fixpoint, put each c3s into C.


relationships and discrete values.

* Lopes, D., Hammoudi, S., Abdelouahab, Z.: Schema Matching in the Context of Model Driven Engineering: From Theory to Practice. In Proceedings of SCSS 2005, pp. 219-227 (2005).

We need more than metamodel

matching based on dictionaries,


relationships and discrete values.



• Another Algorithm for Metamodel Matching

– Our contribution is:

• another algorithm based on structural comparison between a class and its neighbor classes.

– Our algorithm proposal is an extension and

enhancement of the algorithm presented by U. Chukmol

14

enhancement of the algorithm presented by U. Chukmol

et al *.

– We adapt and extend this algorithm to handle

metamodel matching.

* Chuckmol, U., Rifaiem, R., Benharkat, N.: EXSMAL: EDI/XML Semi-Automatic Schema Matching Algorithm. In: Proc. the Seventh IEEE Int. Conference on E-Commerce Technology, pp. 422-425 (2005).




– The similarity function between two classes c1 and c2 is

given by:

( )

coefStructccstructSim

coefBaseccbasicSimccsimilarity

*),(

),(, 2121 +∗=

15

coefStructccstructSim *),( 21

1 and

,10

,1 0

=+

≤≤

≤≤

coefStructcoefBase

coefStruct

coefBasewhere




– The value of similarity(c1,c2) is compared to a threshold

value in the range [0,1].

– If a similarity is greater than threshold value, then c1

and c2 are correspondent.

16

and c2 are correspondent.

– Threshold value is an important point to take a

decision:

• If the value is low, many elements will be considered correspondent in a wrong way (false positive).

• If the value is high, many elements will not be considered as correspondent (false negative).



• Another Algorithm for Metamodel Matching– The structural neighbors of a class C is constituted of:

<ancestor(C), sibling(C),immediateChild(C),leaf(C)>

17




– The structural similarity is obtained from partial

similarities:

• ancestorSimClass(c1,c2)

• siblingSimClass(c1,c2)

Each function populates an array M with dimension

m1 x m2.

where:

18

• immediateChildSimClass(c1,c2)

• leafSimClass(c1,c2)

where:

m1 is the size of the set of

classes related to c1.

m2 is the size of the set of

classes related to c2.



• Another Algorithm for Metamodel Matching– Algorithm for determining similarity between pairs of

ancestors, ancestorSimClass(c1,c2).

for i = 0 until size of (ancestor(c1)) do

for j = 0 until size of (ancestor(c2)) do

M[i][j] <- basicSim (ancestor(c1).get(i),

ancestor(c2).get(j));

19

ancestor(c2).get(j));

endfor;

endfor;

Only based on taxonomies and

dictionaries.

Means that ancestor class of

class c1 (R0) has a similarity

value of 0.7 in relationship to

ancestor class of c2 (C1).



• Another Algorithm for Metamodel Matching– After the creation of M, each function of partial similarity

invokes a function agg.

float agg(float m[][], float thr, float lastvl) {

float average = avg(m, lastvl);

float standardDeviation = sd(m, lastvl);

20

float vc = (float) (standardDeviation / average);

float result = 0;

if (thr >= vc) { return average;}

else {

float lastvl = average * (1 - thr);

result = agg (m, thr, lastvl);

}

return result;

}








21


float result = 0;


else {



}

return result;

}

( ))2(*)1(

]][[

)1(

1

)2(

1

cancestorcancestor

jiM

avg

cancestor

i

cancestor

j∑ ∑=

=

=








22


float result = 0;


else {



}

return result;

}

( )

)2(*)1(

]][[)1(

1

)2(

1

2

cancestorcancestor

avgjiMsd

cancestor

i

cancestor

j∑ ∑= =−

=








23


float result = 0;


else {



}

return result;

}

average

viationstandardDeeficientvarianceCo =








If threshold ≥ vc, the variance coefficient is below the established limit, then the

average can be considered a reliable value, and the function agg returns an

average value.

If threshold < vc, the function agg excludes the values of M that are below avg*(1-

threshold), because these values are contributing to dispersion of the average.

After, agg calculates new measures. This procedure continue until threshold ≥ vc.

24


float result = 0;


else {



}

return result;

}



• Another Algorithm for Metamodel Matching– Function structSim(c1,c2) is given by:

coefimmCccsshildSimClaimmediateC

coefSibccClasssiblingSim

coefAncccmClassancestorSiccstructSim

*)2,1(

*)2,1(

*)2,1(),( 21

+

+

+=

25

coefLeafccssleafSimCla

coefimmCccsshildSimClaimmediateC

*)2,1(

*)2,1( +

1=+++ coefLeafcoefimmCcoefSibcoefAnc

where


Agenda

• Introduction– Problem

– Motivation



• Extending and Adapting SAMT4MDE

26


– Prototyping

– Tests

• Conclusions


Extending and Adapting SAMT4MDE

• Modeling: Semi-Automatic Tool for MDE

ValidateAction GenerateLangAction MatchAction

1

+matchClasses(in classMa, in classMb)

+matchDataTypes(in dTMa, in dTMb)

+matchEnum(in enumMa, in enumMb)

+basicSimClass(in classA, in classB)

+structSimClass(in classA, in classB)

OptmizedMatch

27

+setMouseListener(in mouseControl)

+makeContributions()

-adapterEditingDomain

MappingTreeViewer

11 1 1

1

1

+match()

+init(in pckA, in pckB)

«interface»

ITFMatchEngine+matchClasses(in classMa, in classMb)

+matchDataTypes(in dTMa, in dTMb)

+matchEnum(in enumMa, in enumMb)

+phiClass(in classA, in classB)

+phiEnum(in enumA, in enumB)

+phiDataType(in dTA, in dTB)

Match



• Prototyping

28



• Tests

29



• Tests

30

Results matching UML metamodel and Java

metamodel

Similarity = 0.68

Precision = 0.84

Recall = 0.90

F-Measure = 0.87

Overall = 0.73

threshold=0.6


Agenda


– Problem

– Motivation



31



– Prototyping

– Tests

• Conclusions


Conclusion

• Contribution

– An algorithm for metamodel matching based

on structural similarities.

– The implementation in a tool capable of semi-

automatically creating mapping specifications.

32

• Future works

– Minimize the errors in the matching (i.e. false

positives and false negatives).

– Introduce semantic analysis and machine

learning.


Thank you very much for your attention!

Questions?

AcknowledgementsThis research work is supported by

33

Conselho Nacional de PesquisaFundação de Amparo à Pesquisa e ao Desenvolvimento

Científico e Tecnológico do Maranhão

-

Governo do Estado do Maranhão

Contacts – Denivaldo Lopes (e-mail: [email protected] )

– José de Sousa Jr (e-mail: [email protected])

– Daniela Barreiro Claro (e-mail: [email protected])

– Zair Abdelouahab (e-mail: [email protected])

A Step Forward in Semi-automatic Metamodel Matching: Algorithms and Tool

Documents