Model Transformation By Example 1

1
Model Transformation By Example 1Automating
Model Transformation by Example Using Inductive
Logic Programming 2

• Daniel Varro and Zoltan Balogh
• Budapest Univ. of Technology and Economics
• 1 _at_ ACM/IEEE MoDELS 2006
• 2 _at_ ACM Symposium on Applied Computing 2007

2
Contents

• Research Issue and Objective
• Related Work
• Comparison with MTBE
• The Proposed Approach
• Extension Inductive Logic Programming
• Conclusion

3
Research Issue

• Various model transformation approaches have been
  proposed
• Transformation rule languages such as ATL, QVT,
• All approaches defines transformation rules at
  metamodel (M2) layer
• E.g., An UML Class is mapped into an ER Entity,
• ? It is not easy to define transformation rules
  at M2 layer
• Developers need to learn metamodels
• How M1 elements are represented as instances of
  M2 elements
• Developers need to learn a language to define
  transformation rules
• These languages assume users knows metamodels
  very well

4

• Transformation between UML class diagram and ER
  diagram

5
Research Objective

• Model Transformation By-Example
• Allows defining inter-model mappings on M1 layer
• Developers do not need to consider the M2 layer.
  They can work at the M1 layer
• A large part of transformation design is
  generated semi-automatically
• An initial set of transformation rules
  (automatically generated) can be refined
  iteratively by providing additional source-target
  model pairs

6
Related Work

• Differences from Towards Model
  TransformationGeneration By-Example (MTBE) in
  HICSS 2007
• All elements in source and target models has
  one-to-one relationship
• No deletion, no addition
• Elements with a certain notation on M1 layer in a
  source model is always mapped into a
  corresponding type in a target model
• E.g., Class in UML is always mapped into Entity
  in ER, Classs attribute in UML is mapped into
  Attibute in ER, Associations role in UML is
  mapped into Role in ER,
• The proposed approach in this paper allows
• Elements addition / deletion
• Conditions on M1 layer
• E.g., A class with no parent is mapped into XXX,
  but a class with a parent is mapped into YYY

7
Model Transformation by Example

• Steps in the proposed approach
• Setting up an initial mapping model
• Define source and target models, and mappings
  between elements by hand
• Creation of target nodes
• Derive model transformation rules for each
  mapping
• Context Analysis examine contexts of all source
  and target nodes
• Generating transformation rules
• Interconnection of target nodes
• Derive links between target nodes based on the
  connectivity of the source and target nodes
• Iterative refinement
• Refining the derived transformation rules

8
Initial Mapping Model

• Define source and target models, and mappings
• E.g., UML class diagram and ER diagram

9
(No Transcript)
10
Context Analysis of Source Nodes

• Elements of a certain type in a source model with
  a certain condition (precondition) are mapped
  into elements in a target model
• The Product class is mapped while the Book class
  is unmapped to the Table
• In order identify a precondition, the proposed
  approach examines all mappings of a certain type
  one by one and determine the 1-context of the
  source element
• 1-context a condition to identify elements to be
  mapped
• The approach considers the existence and
  non-existence of incoming and outgoing edges
• Check all incoming and outgoing edges defined in
  a metamodel

11

• Customer has an incoming src edge, an incoming
  parent edge, but no incoming dst edge, no
  incoming type edge, no outgoing attrs edge, and
  no outgoing parent edge
• Joint 1-context The next step is to merge all
  individual 1-contexts
• If an edge of a certain type is present in all
  individual 1-contexts, then it becomes a must
  edge in the joint 1-context
• Forbidden edges and may edges are generated in
  the same way

12

• Checking the 1-context for unmapped elements
• Check all unmapped source nodes to identify a
  negative condition
• A mapping is not applied when a corresponding
  negative condition met
• E.g., VIPCustomer, NormalCustomer, CD and Book
  are not mapped into Entity

13
Context Analysis of Target Nodes

• The context of target nodes aims at identifying
  the postconditions of the transformation rules

• Unmapped nodes which contained by a target node
  are considered as a context
• Contained relationships are defined in a
  metamodel
• E.g., Table contains Column and FKey
• Since CustFavourite and ProdAppendix have their
  own mapping, they are not included in a context

14
Generating Transformation Rules

• Transformation rules are derived in the form of
  graph transformation rules
• Consists of a left-hand side graph (LHS), a
  right-hand side graph (RHS), and a negative
  application condition graph (NAC)
• LHS and NAC are collectively called the
  precondition (PRE)
• Deriving a graph transformation rule from joint
  contexts
• Joint source context ? LHS
• Negative condition ? NAC
• Joint target context ? RHS
• This paper does not mention (does not
  consider?) how to map the name of nodes

15
Joint Source Context
Joint Target Context
Transformation Rule
16
Extending the mapping

• In the previous rule, a Class that has subclasses
  but no attributes is mapped into an Entity
• By providing a new example, a transformation rule
  can be refined

17
Interconnection of Target Nodes

• In the previous step, each type is investigated
  separately. The next step is to investigate the
  relationship between mappings (target nodes).
• Check all combinations of mappings
• Two source nodes identified by different mappings
  are connected with a path of edges via unmapped
  source nodes

• Target nodes generated by r6 and r2 should have a
  connection(s)

18

• Find a connection between target nodes

19

• Generate a rule for creating a cref
• A LHS is a joint context

20
Extension Inductive Logic Programming

• The quality of source and target contexts depends
  on the quality of source and target models
• If mappings between source and target models
  cover many transformation patterns, generated
  transformation rule will be good (conditions
  reflect users intention well)
• New way to analyze context is required to find
  more conditions from a relatively small model
• Not only checking incoming/outgoing edges, but
  also considering other elements
• E.g., consider not only relations with one-hop
  neighbors, but also multi-hop neighbors
• ? Inductive Logic Programming (ILP)

21
Whats the ILP?

• ILP is a subfield of machine learning
• Given the known background knowledge and a set of
  positive/negative facts, an ILP system will
  derive hypotheses which entails all the positive
  and none of the negative facts
• Background knowledge positive facts negative
  facts ? hypotheses

22
Background knowledge
Hypothesis
Positive Facts
Negative Facts

• The authors use Aleph
• http//web.comlab.ox.ac.uk/oucl/research/areas/mac
  hlearn/Aleph/
• Accepts inputs in Prolog

23
Context Analysis

• Background knowledge
• from Class diagram
• Facts
• from mappings

24

• ILP can identify negative conditions by means of
  constraint learning
• Constraint learning recording new constraints
  whenever an inconsistency is found
• E.g., Only top-level classes are transformed into
  a table

25
Connectivity Analysis

• To analyze connectivity among target nodes,
  positive and negative facts are derived from an
  edge in the target model

26
Conclusion

• The authors propose Model Transformation by
  Example to allow the definition of inter-model
  mappings between concrete models on the M1 layer
• ILP helps find pre/post conditions of rules
• No evaluation, but the authors say it works well
  and exceeds our expectations