The document outlines an approach for scalable quality-driven semantic web service composition. It discusses semantic links between services' input and output parameters, and defines quality criteria for these links and services. A composition is modeled as a statechart with states as services and transitions as semantic links. Quality of a link is measured by its common description rate and matching quality, while services have associated non-functional properties for quality of service. The approach aims to select compositions based on functional constraints from semantic links and quality of service constraints.
Scalable Quality Driven Semantic Web Service Composition
1. Introduction Related Work Background Quality Model A Scalable Approach Experimentation Conclusion
Towards Scalability of Quality Driven
Semantic Web Service Composition
Freddy Lécué1 Nikolay Mehandjiev1
firstname.lastname@manchester.ac.uk
1 The University of Manchester
Booth Street East, Manchester, UK
IEEE 7th International Conference on Web Services
(ICWS 2009)
July 6th - 10th , 2009
Los Angeles, CA, USA
2. Introduction Related Work Background Quality Model A Scalable Approach Experimentation Conclusion
Outline
1 Introduction
2 Related Work
3 Background
4 Quality Model
5 A Scalable Approach
6 Experimentation
7 Conclusion
3. Introduction Related Work Background Quality Model A Scalable Approach Experimentation Conclusion
Big Picture
(Automation of) Web service composition in the Semantic Web.
Rather than Optimization (NP Hard!), we address ...
Scalability of quality driven semantic service composition:
by selecting compositions based on:
functional constraints;
and QoS constraints.
T2 T3 T6
s2 s3 s6
T1 T5 T8
s1 s5 s8
T4 T7
s4 s7
4. Introduction Related Work Background Quality Model A Scalable Approach Experimentation Conclusion
Most of approaches focus on optimization wrt:
QoS or functional parameters.
L. Zeng, B. Benatallah et al.
Quality Driven Web Services Composition.
In WWW, pages 411–421, 2003.
F. Lecue, A. Leger and A. Delteil
Optimizing Causal Link Based Web Service Composition.
In ECAI, pages 45–49, 2008.
Our work focuses on composition selection wrt:
both latter criteria!
Semantics based T2 T3 T6
Selection s3 s6
T1 s1, s2, s3, ...
2 2 2
T5 T8
s1 QoS based s5 s8
Selection
T4 T7
s4 s7
5. Introduction Related Work Background Quality Model A Scalable Approach Experimentation Conclusion
Semantic Web Service in a Nutshell
Parameters (i.e., Input and Output) of Web services in
semantic Web are concepts referred to in an ontology T :
WSDL-S, SA-WSDL (W3C Proposed Recommendation);
OWL-S profile level;
WSMO capability level.
6. Introduction Related Work Background Quality Model A Scalable Approach Experimentation Conclusion
Web Service Composition and its Semantic Links
Semantic Link: Semantic connection between services;
... more particulary between Output and Input parameters;
... denoted by sly ,x and valued by SimT (Out_sy , In_sx );
S y Output S x Input
Parameters Parameters
S y Input S x Output
Parameters Parameters
Out_sy In_sx
Web service: sx
Web service: sy
SimT is reduced to the five matchmaking functions
[M.Paolucci et al. ISWC’02, Li and Horrocks WWW’03]:
Exact i.e., T |= Out_sy ≡ In_sx ;
PlugIn i.e., T |= Out_sy In_sx ;
Subsume i.e., T |= In_sx Out_sy ;
Intersection i.e., T |= Out_sy In_sx ⊥;
Disjoint i.e., T |= Out_sy In_sx ⊥;
7. Introduction Related Work Background Quality Model A Scalable Approach Experimentation Conclusion
Web Service Composition and its Semantic Links
Semantic Link: Semantic connection between services;
... more particulary between Output and Input parameters;
... denoted by sly ,x and valued by SimT (Out_sy , In_sx );
S y Output S x Input
Parameters Parameters
S y Input S x Output
Parameters Parameters
Out_sy In_sx
Semantic connection: Sim Web service: sx
T
Web service: sy
SimT is reduced to the five matchmaking functions
[M.Paolucci et al. ISWC’02, Li and Horrocks WWW’03]:
Exact i.e., T |= Out_sy ≡ In_sx ;
PlugIn i.e., T |= Out_sy In_sx ;
Subsume i.e., T |= In_sx Out_sy ;
Intersection i.e., T |= Out_sy In_sx ⊥;
Disjoint i.e., T |= Out_sy In_sx ⊥;
8. Introduction Related Work Background Quality Model A Scalable Approach Experimentation Conclusion
Web Service Composition and its Semantic Links
Semantic Link: Semantic connection between services;
... more particulary between Output and Input parameters;
... denoted by sly ,x and valued by SimT (Out_sy , In_sx );
S y Output S x Input
Parameters Parameters
S y Input S x Output
Parameters Parameters
Out_sy In_sx
Semantic connection: Sim Web service: sx
T
Web service: sy
SimT is reduced to the five matchmaking functions
[M.Paolucci et al. ISWC’02, Li and Horrocks WWW’03]:
Exact i.e., T |= Out_sy ≡ In_sx ;
PlugIn i.e., T |= Out_sy In_sx ;
Subsume i.e., T |= In_sx Out_sy ;
Intersection i.e., T |= Out_sy In_sx ⊥;
Disjoint i.e., T |= Out_sy In_sx ⊥;
9. Introduction Related Work Background Quality Model A Scalable Approach Experimentation Conclusion
Web Service Composition and its Semantic Links
Semantic Link: Semantic connection between services;
... more particulary between Output and Input parameters;
... denoted by sly ,x and valued by SimT (Out_sy , In_sx );
S y Output S x Input
Parameters Parameters
S y Input S x Output
Parameters Parameters
NetWorkConnection
NetWorkConnection
Semantic connection: Sim Web service: sx
T
Web service: sy
SimT is reduced to the five matchmaking functions
[M.Paolucci et al. ISWC’02, Li and Horrocks WWW’03]:
Exact i.e., T |= Out_sy ≡ In_sx ;
PlugIn i.e., T |= Out_sy In_sx ;
Subsume i.e., T |= In_sx Out_sy ;
Intersection i.e., T |= Out_sy In_sx ⊥;
Disjoint i.e., T |= Out_sy In_sx ⊥;
10. Introduction Related Work Background Quality Model A Scalable Approach Experimentation Conclusion
Web Service Composition and its Semantic Links
Semantic Link: Semantic connection between services;
... more particulary between Output and Input parameters;
... denoted by sly ,x and valued by SimT (Out_sy , In_sx );
S y Output S x Input
Parameters Parameters
S y Input S x Output
Parameters Parameters
NetWorkConnection
SlowNetWorkConnection
Semantic connection: Sim Web service: sx
T
Web service: sy
SimT is reduced to the five matchmaking functions
[M.Paolucci et al. ISWC’02, Li and Horrocks WWW’03]:
Exact i.e., T |= Out_sy ≡ In_sx ;
PlugIn i.e., T |= Out_sy In_sx ;
Subsume i.e., T |= In_sx Out_sy ;
Intersection i.e., T |= Out_sy In_sx ⊥;
Disjoint i.e., T |= Out_sy In_sx ⊥;
11. Introduction Related Work Background Quality Model A Scalable Approach Experimentation Conclusion
Web Service Composition and its Semantic Links
Semantic Link: Semantic connection between services;
... more particulary between Output and Input parameters;
... denoted by sly ,x and valued by SimT (Out_sy , In_sx );
S y Output S x Input
Parameters Parameters
S y Input S x Output
Parameters Parameters
SlowNetWorkConnection
NetWorkConnection
Semantic connection: Sim Web service: sx
T
Web service: sy
SimT is reduced to the five matchmaking functions
[M.Paolucci et al. ISWC’02, Li and Horrocks WWW’03]:
Exact i.e., T |= Out_sy ≡ In_sx ;
PlugIn i.e., T |= Out_sy In_sx ;
Subsume i.e., T |= In_sx Out_sy ;
Intersection i.e., T |= Out_sy In_sx ⊥;
Disjoint i.e., T |= Out_sy In_sx ⊥;
12. Introduction Related Work Background Quality Model A Scalable Approach Experimentation Conclusion
Web Service Composition and its Semantic Links
Semantic Link: Semantic connection between services;
... more particulary between Output and Input parameters;
... denoted by sly ,x and valued by SimT (Out_sy , In_sx );
S y Output S x Input
Parameters Parameters
S y Input S x Output
Parameters Parameters
IPAddress
VoIPId
Semantic connection: Sim Web service: sx
T
Web service: sy
SimT is reduced to the five matchmaking functions
[M.Paolucci et al. ISWC’02, Li and Horrocks WWW’03]:
Exact i.e., T |= Out_sy ≡ In_sx ;
PlugIn i.e., T |= Out_sy In_sx ;
Subsume i.e., T |= In_sx Out_sy ;
Intersection i.e., T |= Out_sy In_sx ⊥;
Disjoint i.e., T |= Out_sy In_sx ⊥;
13. Introduction Related Work Background Quality Model A Scalable Approach Experimentation Conclusion
Web Service Composition and its Semantic Links
Semantic Link: Semantic connection between services;
... more particulary between Output and Input parameters;
... denoted by sly ,x and valued by SimT (Out_sy , In_sx );
S y Output S x Input
Parameters Parameters
S y Input S x Output
Parameters Parameters
Address
NetWorkConnection
Semantic connection: Sim Web service: sx
T
Web service: sy
SimT is reduced to the five matchmaking functions
[M.Paolucci et al. ISWC’02, Li and Horrocks WWW’03]:
Exact i.e., T |= Out_sy ≡ In_sx ;
PlugIn i.e., T |= Out_sy In_sx ;
Subsume i.e., T |= In_sx Out_sy ;
Intersection i.e., T |= Out_sy In_sx ⊥;
Disjoint i.e., T |= Out_sy In_sx ⊥;
14. Introduction Related Work Background Quality Model A Scalable Approach Experimentation Conclusion
Extra & Common Description in Semantic Links
Definition (Concept Abduction)
Let L be a DL, Out_si , In_sj be two concepts in L, and T be a
set of axioms in L. A Concept Abduction Problem (CAP),
denoted as L, Out_si , In_sj , T is finding a concept H ∈ L
such that T |= Out_si H In_sj .
The Extra Description H represents what is underspecified in
Out_si in order to completely satisfy In_sj ;
⇒ Explain why Out_si and In_sj can not be chained by a
robust semantic link.
The Common Description CD lcs(Out_si , In_sj ) refers to
information required by In_sj and effectively provided by
Out_si .
15. Introduction Related Work Background Quality Model A Scalable Approach Experimentation Conclusion
Extra & Common Description in Semantic Links
Definition (Concept Abduction)
Let L be a DL, Out_si , In_sj be two concepts in L, and T be a
set of axioms in L. A Concept Abduction Problem (CAP),
denoted as L, Out_si , In_sj , T is finding a concept H ∈ L
such that T |= Out_si H In_sj .
e.g., in case of a semantic link valued by the Subsume
matching type.
S y Output S x Input
Parameters Parameters
S y Input S x Output
Parameters Parameters
Web service: sx
Web service: sy
16. Introduction Related Work Background Quality Model A Scalable Approach Experimentation Conclusion
Extra & Common Description in Semantic Links
Definition (Concept Abduction)
Let L be a DL, Out_si , In_sj be two concepts in L, and T be a
set of axioms in L. A Concept Abduction Problem (CAP),
denoted as L, Out_si , In_sj , T is finding a concept H ∈ L
such that T |= Out_si H In_sj .
e.g., in case of a semantic link valued by the Subsume
matching type.
S y Output S x Input
Parameters Parameters
S y Input S x Output
Parameters Parameters
Web service: sx
Web service: sy
17. Introduction Related Work Background Quality Model A Scalable Approach Experimentation Conclusion
Composition Model
Process Model as a Statechart
Its states refer to services;
Its transitions are labelled with semantic links;
with basic composition constructs such as Sequence,
conditional branching (i.e., OR-Branching), concurrent
threads (i.e., AND-Branching).
T2 T3 T6 Legend
Slow s2 s3 s6 Semantic Link sl
1
1
sl1,2 Network 1
sl2,3 1
sl5,6 1
sl6,8
Connection sl3,5 Input Parameter
T1 T5 T8
Network Sequence AND Output Parameter
s1 s5 Branching s8
Connection
OR-Branching T: Task
1 1 1
T4 sl4,5 sl5,7 sl7,8
1
sl1,4 T7 s: Service
s4 s7
18. Introduction Related Work Background Quality Model A Scalable Approach Experimentation Conclusion
Quality Criteria for Semantic Links & Services
q(sli,j ) for Elementary Semantic Links sli,j
Common Description rate qcd ∈ (0, 1]:
|lcs(Out_si , In_sj )|
qcd (sli,j ) =
|H∈ L,Out_si ,In_sj ,T | + |lcs(Out_si , In_sj )|
Matching Quality qm ∈ (0, 1], valued by SimT (Out_si , In_sj )
3 1 1
(Exact: 1, PlugIn: 4
, Subsume: 2
, Intersection: 4
).
19. Introduction Related Work Background Quality Model A Scalable Approach Experimentation Conclusion
Quality Criteria for Semantic Links & Services
q(sli,j ) for Elementary Semantic Links sli,j
Common Description rate qcd ∈ (0, 1]:
|lcs(Out_si , In_sj )|
qcd (sli,j ) =
|H∈ L,Out_si ,In_sj ,T | + |lcs(Out_si , In_sj )|
Matching Quality qm ∈ (0, 1], valued by SimT (Out_si , In_sj )
3 1 1
(Exact: 1, PlugIn: 4
, Subsume: 2
, Intersection: 4
).
S y Output (Subsume i.e., 1 ) S x Input
_
Parameters Parameters
S y Input 2 S x Output
Parameters Parameters
SlowNetWorkConnection
NetWorkConnection
Semantic connection: Sim Web service: sx
T
Web service: sy
20. Introduction Related Work Background Quality Model A Scalable Approach Experimentation Conclusion
Quality Criteria for Semantic Links & Services
q(sli,j ) for Elementary Semantic Links sli,j
Common Description rate qcd ∈ (0, 1]:
|lcs(Out_si , In_sj )|
qcd (sli,j ) =
|H∈ L,Out_si ,In_sj ,T | + |lcs(Out_si , In_sj )|
Matching Quality qm ∈ (0, 1], valued by SimT (Out_si , In_sj )
3 1 1
(Exact: 1, PlugIn: 4
, Subsume: 2
, Intersection: 4
).
q(si ) for Elementary Services si
Execution Price qpr ∈ +;
Response Time qt ∈ +.
21. Introduction Related Work Background Quality Model A Scalable Approach Experimentation Conclusion
Quality Criteria for Semantic Links & Services
q(sli,j ) for Elementary Semantic Links sli,j
Common Description rate qcd ∈ (0, 1]:
|lcs(Out_si , In_sj )|
qcd (sli,j ) =
|H∈ L,Out_si ,In_sj ,T | + |lcs(Out_si , In_sj )|
Matching Quality qm ∈ (0, 1], valued by SimT (Out_si , In_sj )
3 1 1
(Exact: 1, PlugIn: 4
, Subsume: 2
, Intersection: 4
).
q(si ) for Elementary Services si
Execution Price qpr ∈ +;
Response Time qt ∈ +.
QoS-extended quality vector of a semantic link sli,j
∗ .
q (sli,j ) = (q(si ), q(sli,j ), q(sj ))
22. Introduction Related Work Background Quality Model A Scalable Approach Experimentation Conclusion
Quality Criteria for Composition
Quality Aggregation Rules for Compositions
Quality Criterion
Composition
Semantic Non Functional
Construct
Qcd Qm Qt Qpr
Sequential/ 1 s qt (s)
|sl| sl qcd (sl) sl qm (sl) s qpr (s)
AND- Branching maxs qt (s)
OR-Branching sl qcd (sl).psl sl qm (sl).psl s qt (s).ps s qpr (s).ps
T2 T3 T6 Legend
Slow s
2 s3 s6 Semantic Link sl
1
1
sl1,2 Network 1
sl2,3 1
sl5,6 1
sl6,8
Connection sl3,5 Input Parameter
T1 T5 T8
Network Sequence AND Output Parameter
s1 s5 Branching s8
Connection
OR-Branching T: Task
1 1 1
T4 sl4,5 sl5,7 sl7,8
1
sl1,4 T7 s: Service
s4 s7
23. Introduction Related Work Background Quality Model A Scalable Approach Experimentation Conclusion
Quality Criteria for Composition
Quality Aggregation Rules for Compositions
Quality Criterion
Composition
Semantic Non Functional
Construct
Qcd Qm Qt Qpr
Sequential/ 1 s qt (s)
|sl| sl qcd (sl) sl qm (sl) s qpr (s)
AND- Branching maxs qt (s)
OR-Branching sl qcd (sl).psl sl qm (sl).psl s qt (s).ps s qpr (s).ps
T2 T3 T6 Legend
Slow s
2 s3 s6 Semantic Link sl
1
1
sl1,2 Network 1
sl2,3 1
sl5,6 1
sl6,8
Connection sl3,5 Input Parameter
T1 T5 T8
s1 Network
s5 s8 Output Parameter
Connection
T: Task
Sequence s: Service
24. Introduction Related Work Background Quality Model A Scalable Approach Experimentation Conclusion
Quality Criteria for Composition
Quality Aggregation Rules for Compositions
Quality Criterion
Composition
Semantic Non Functional
Construct
Qcd Qm Qt Qpr
Sequential/ 1 s qt (s)
|sl| sl qcd (sl) sl qm (sl) s qpr (s)
AND- Branching maxs qt (s)
OR-Branching sl qcd (sl).psl sl qm (sl).psl s qt (s).ps s qpr (s).ps
T6 Legend
s6 Semantic Link sl
1
sl5,6
Input Parameter
T5
AND Output Parameter
s5 Branching
T: Task
1
sl5,7
T7 s: Service
s7
25. Introduction Related Work Background Quality Model A Scalable Approach Experimentation Conclusion
Quality Criteria for Composition
Quality Aggregation Rules for Compositions
Quality Criterion
Composition
Semantic Non Functional
Construct
Qcd Qm Qt Qpr
Sequential/ 1 s qt (s)
|sl| sl qcd (sl) sl qm (sl) s qpr (s)
AND- Branching maxs qt (s)
OR-Branching sl qcd (sl).psl sl qm (sl).psl s qt (s).ps s qpr (s).ps
T2 Legend
Slow s
2 Semantic Link sl
1
sl1,2 Network
Connection Input Parameter
T1
s1 Network Output Parameter
Connection
OR-Branching T: Task
1
sl1,4 T4 s: Service
s4
26. Introduction Related Work Background Quality Model A Scalable Approach Experimentation Conclusion
Quality Criteria for Composition
Quality Aggregation Rules for Compositions
Quality Criterion
Composition
Semantic Non Functional
Construct
Qcd Qm Qt Qpr
Sequential/ 1 s qt (s)
|sl| sl qcd (sl) sl qm (sl) s qpr (s)
AND- Branching maxs qt (s)
OR-Branching sl qcd (sl).psl sl qm (sl).psl s qt (s).ps s qpr (s).ps
A Quality Vector for Web Service Composition
“A” way to differentiate compositions:
.
Q(c) = (Qcd (c), Qm (c), Qt (c), Qpr (c))
27. Introduction Related Work Background Quality Model A Scalable Approach Experimentation Conclusion
Web Service Composition Driven CSP
CSP Formalization
Formalization as a triple (T , D, C):
T is the set of tasks (variables) {T1 , T2 , ..., Tn };
D is the set of domains {D1 , D2 , ..., Dn } i.e., services;
C is the set of constraints i.e., local CL and global CG .
1 A +
e.g., A
qcd (sli,j ) ≥ v , v ∈ [0, 1] qpr (Ti ) ≤ v , v ∈
|sli,j | A Ti
sli,j
Main Goal to Achieve
An assignment (si , Ti )1≤i≤n i.e., (service, task)
with si,1≤i≤n ∈ Di,1≤i≤n ;
which satisfies all the constraints C.
28. Introduction Related Work Background Quality Model A Scalable Approach Experimentation Conclusion
A Stochastic Search Method (1)
Principles
Sacrificing completness (i.e., all solutions) for speed;
Based on a simple idea: computing “a single” solution.
Our Approach
Adaptation of the Hill Climbing algorithm.
→ Appropriate for a large number of services.
S. Russell and P. Norvig.
Artificial Intelligence: A Modern Approach.
Ed. Prentice-Hall, 1995.
Computational Complexity
CSP based search methods: Exponential!
Stochastic search methods (e.g., Hill Climbing) scale better!
29. Introduction Related Work Background Quality Model A Scalable Approach Experimentation Conclusion
A Stochastic Search Method (2)
Requirements
An evaluation function f for each composition c:
ˆ ˆ
ωcd Qcd (c) + ωm Qm (c)
f (c) =
ˆ ˆ
ωpr Qpr (c) + ωt Qt (c)
An adjacency function: c1 and c2 are adjacent to each
other if they differ in exactly one assignment (s, T ).
Algorithm in Details
1) Let’s start with a random composition cfinal .
2) f -Evaluation of all ci,1≤i≤n adjacent to cfinal .
If ∃i such that f (cfinal ) ≤ f (ci ) then f (cfinal ) ← f (ci ).
3) Iteration until all constraints are satisfied by cfinal .
4) If no solution, constraints relaxing.
30. Introduction Related Work Background Quality Model A Scalable Approach Experimentation Conclusion
Evolution of Constraints Satisfaction
The more tasks, services the more time consuming!
Evolution of Composition Quality
Optimal composition: High Time consuming!
Compositions that satisfy constraints: More scalable!
Search Process vs. DL Reasoning (|T | > 100, |s| > 350)
DL reasoning is the most time consuming process!
Large number of potential semantic links.
Critical complexity of DL abduction.
Vs. State-of-the-art Approaches (T = 300 |s| > 280)
Adoption of stochastic search method for large domains!
No exponential search required.
31. Introduction Related Work Background Quality Model A Scalable Approach Experimentation Conclusion
Evolution of Constraints Satisfaction
The more tasks, services the more time consuming!
Evolution of Composition Quality
Optimal composition: High Time consuming!
Compositions that satisfy constraints: More scalable!
Search Process vs. DL Reasoning (|T | > 100, |s| > 350)
DL reasoning is the most time consuming process!
Large number of potential semantic links.
Critical complexity of DL abduction.
Vs. State-of-the-art Approaches (T = 300 |s| > 280)
Adoption of stochastic search method for large domains!
No exponential search required.
32. Introduction Related Work Background Quality Model A Scalable Approach Experimentation Conclusion
Evolution of Constraints Satisfaction
The more tasks, services the more time consuming!
Evolution of Composition Quality
Optimal composition: High Time consuming!
Compositions that satisfy constraints: More scalable!
Search Process vs. DL Reasoning (|T | > 100, |s| > 350)
DL reasoning is the most time consuming process!
Large number of potential semantic links.
Critical complexity of DL abduction.
Vs. State-of-the-art Approaches (T = 300 |s| > 280)
Adoption of stochastic search method for large domains!
No exponential search required.
33. Introduction Related Work Background Quality Model A Scalable Approach Experimentation Conclusion
Evolution of Constraints Satisfaction
The more tasks, services the more time consuming!
Evolution of Composition Quality
Optimal composition: High Time consuming!
Compositions that satisfy constraints: More scalable!
Search Process vs. DL Reasoning (|T | > 100, |s| > 350)
DL reasoning is the most time consuming process!
Large number of potential semantic links.
Critical complexity of DL abduction.
Vs. State-of-the-art Approaches (T = 300 |s| > 280)
Adoption of stochastic search method for large domains!
No exponential search required.
34. Introduction Related Work Background Quality Model A Scalable Approach Experimentation Conclusion
Quality-driven Semantic Web Service Composition
[Theoretical]
A general and extensible model to evaluate compositions;
Scalabilty: A solution rather than the most optimal.
CSP formalization.
Adaptation of a stochastic search method.
[Experimental]
Good computation costs despite the off-line DL reasoning.
Future Work
Extension with more composition constructs;
Considering a finer abduction operator;
Dynamic Distribution of the CSP on different peers;
Focusing on a process that reduces the number of
semantic links: Macro composition of Web services.
35. Introduction Related Work Background Quality Model A Scalable Approach Experimentation Conclusion
Question?