Académique Documents
Professionnel Documents
Culture Documents
2, FEBRUARY 2010
Abstract—In this paper, we describe and evaluate a Web service discovery framework using OWL-S advertisements, combined with
the distinction between service and Web service of the WSMO Discovery Framework. More specifically, we follow the Web service
discovery model, which is based on abstract and lightweight semantic Web service descriptions, using the Service Profile ontology of
OWL-S. Our goal is to determine fast an initial set of candidate Web services for a specific request. This set can then be used in more
fine-grained discovery approaches, based on richer Web service descriptions. Our Web service matchmaking algorithm extends
object-based matching techniques used in Structural Case-based Reasoning, allowing 1) the retrieval of Web services not only based
on subsumption relationships, but exploiting also the structural information of OWL ontologies and 2) the exploitation of Web services
classification in Profile taxonomies, performing domain-dependent discovery. Furthermore, we describe how the typical paradigm of
Profile input/output annotation with ontology concepts can be extended, allowing ontology roles to be considered as well. We have
implemented our framework in the OWLS-SLR system, which we extensively evaluate and compare to the OWLS-MX matchmaker.
Index Terms—Web service discovery, abstract descriptions, OWL-S profile, structural information, role-oriented matchmaking.
1 INTRODUCTION
Authorized licensd use limted to: Imperial Coleg Lond. Downlade on June 07,21 at 19:3826 UTC from IE Xplore. Restricon aply.
280 IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, VOL. 22, NO. 2, FEBRUARY 2010
are described as direct instances of the Profile concept, the Definition 1. An Object O is a triple hID; C; P i, where ID is
domain knowledge can only be captured through special the unique identifier of the object, C is the object class type,
roles, since all the instances belong to the same concept. In and P is a set that contains attribute-value pairs of the form
our framework, we allow advertisements to be defined in hp; V i, where p is an attribute and V a set of values.
terms of Profile taxonomies, such as the Profile instances of
Table 1, capturing domain knowledge through instance class The domain model is represented as a class hierarchy and
memberships (DL ABox reasoning). the objects are initialized with a single class type and
property-value definitions. For example, let A B denote
2.2.2 Structural Ontological Knowledge that class A is subclass of class B; p 2 AttðAÞ denote that the
The matchmaking that is based only on logic-based attribute p is defined in class A; o:p denote the set of values of
reasoning, such as the CC approach, computes only the object o for property p, that is, the set V , and o 7! A denote
subsumption relationships among the annotation concepts. the class type of object o. Let three classes A; B, and D, where
Therefore, any structural information is ignored, for A B and B D. If o 7! A, then o is also an object of B and
example, sibling relationships that may enhance the D, due to inheritance. Furthermore, let pA and pB be two
discovery, especially in cases where few or no results are attributes, where pA 2 AttðAÞ and pB 2 AttðBÞ. If o 7! A,
initially returned for a request. then both expressions o:pA and o:pB are valid, since
We are motivated by the usefulness of the structural attributes are inherited to subclasses.
information and we introduce in our matchmaking algo- In this way, every object encapsulates domain knowl-
rithm matching techniques that are used in SCBR, a edge, regarding class relationships and property-value
specialized approach to Case-based Reasoning. In SCBR, definitions, which is used for matching cases and queries
the idea is to represent cases according to a domain model through the interclass and intraclass similarity metrics.
[6] that is structured in an object-oriented manner,
including IS-A relationships and inheritance. Each case 3.1 Intraclass Similarity
and query is modeled as an object and the additional The intraclass metric defines the similarity of two objects in
knowledge that stems from the model is used during terms of the values in their common attributes, based on
matching. The object relevance is determined using the two value matching functions: the Vs function for simple
interclass and intraclass metrics. The former is defined over values, for example, integers, strings, etc., and the Vr
the common attributes, whereas the latter is defined upon function for relational values, that is, objects.
the class types of two objects. Let two objects OA ¼ hIDA ; CA ; PA i and OB ¼ hIDB ;
In our work, we extend the intraclass and interclass CB ; PB i and their common attribute p. The partial intraclass
notions to the domain of SWS discovery. The idea is to similarity Sp for the property p is defined as
perform matchmaking on Profile instances represented as
objects, considering the domain ontologies and any Profile Vs ðIDA :p; IDB :pÞ; if p is simple;
Sp ðOA ; OB Þ ¼
taxonomy as the domain models. Vr ðIDA :p; IDB :pÞ; if p is relational:
The overall intraclass similarity of two objects OA and OB
2.2.3 Web Service Descriptions and Ontology Roles
is defined by aggregating their partial similarities.
The semantic tagging of I/O parameters with predefined
concepts has limited expressive power. For example, Definition 2. Let two objects OA ¼ hIDA ; CA ; PA i and OB ¼
consider a Web service advertisement whose one of the hIDB ; CB ; PB i and the set T of their common attributes, that
inputs is annotated with the concept Person that has three is, 8p 2 T ; 9hp; V i 2 PA ^ 9hp; V 0 i 2 PB . The intraclass simi-
roles: SSN, address, and name. In this case, we cannot larity is defined, with respect to an aggregation function , as
determine what the Web service really requires: SSN, name,
Sintra ðOA ; OB Þ ¼ Sp ðOA ; OB Þ:
address, or all of them? On the other hand, the CC approach 8p2T
takes into account roles through restrictions.
In order to overcome this limitation of the SP paradigm, 3.2 Interclass Similarity
we enhance our framework with the ability of a role-based
The interclass metric captures the hierarchical relationship
Web service functional annotation based on the open-world
of two object class types, based on a hierarchical matching
assumption and the classification capabilities of the
function H that denotes the similarity of two objects in terms
DL reasoning paradigm. In this way, we are able to extend
of their class types.
the annotation and discovery procedures of our SP-oriented
framework with ontology roles. Actually, our approach is Definition 3. Let two objects OA ¼ hIDA ; CA ; PA i and OB ¼
an effort to leverage the modeling differences between the hIDB ; CB ; PB i. Their interclass similarity is defined, with
CC and SP paradigms, where the former is concept- respect to a hierarchical matching function H, as
oriented, allowing the full exploitation of the logical
formalism that is used to define concepts, whereas the Sinter ðOA ; OB Þ ¼ HðCA ; CB Þ:
latter is instance-oriented, treating Web service descriptions
The overall similarity of the two objects is defined by
as Profile instances.
aggregating their intraclass and interclass similarities.
Definition 4. Let two objects OA and OB . Their similarity S is
3 SCBR SIMILARITY METRICS defined, with respect to an aggregation function , as
In SCBR, both cases and queries are represented as objects,
enhancing the typical attribute-value representation of the SðOA ; OB Þ ¼ ½Sintra ðOA ; OB Þ; Sinter ðOA ; OB Þ:
traditional CBR with domain knowledge.
Authorized licensd use limted to: Imperial Coleg Lond. Downlade on June 07,21 at 19:3826 UTC from IE Xplore. Restricon aply.
MEDITSKOS AND BASSILIADES: STRUCTURAL AND ROLE-ORIENTED WEB SERVICE DISCOVERY WITH TAXONOMIES IN OWL-S 281
f
4 OWL-S PROFILE METRICS We generalize the A B relation to a set of filters F and
In this section, we describe the DLH and DLR Profile- define that concept A matches concept B, with respect to a
aware similarity metrics, extending the intraclass and filter set F , if and only if there is at least one filter f in F
f
interclass SCBR metrics to an ontology environment, and such that A B, that is,
enhancing them with DL reasoning. First, we introduce the
notion of the object specification for representing advertise- F f
A B , 9f 2 F : A B:
ment and query instances in our framework.
Definition 5. An object specification is a quintuple hI D; C; Definition 6. Let two concepts X and Y . Their DLH similarity is
I ; O; N F i, where ID is the Profile instance identifier, C is the the normalized value to ½0::1 that is defined, with respect to a
set of the most specific concepts to where ID belongs, I and O concept similarity function S and a hierarchical filter set F , as
are the sets of I/O annotation concepts, respectively, and N F
is the set of nonfunctional property-value pairs. F
DLHðX; Y ; F Þ ¼ SðX; Y Þ if X Y ; ð1Þ
We refer to an advertisement instance as an A specifica- 0 otherwise:
tion and to a query instance as a Q specification. In this
way, the Profile instances of Table 1, which are used as We generalize (1) on two sets SA ; SB of concepts as
examples in the rest of the paper using the a1 advertisement P
as a query, are represented as max8A2SA ½DLHðB; A; F Þ
DLHset ðSA ; SB ; F Þ ¼ 8B2SB : ð2Þ
jSB j
Q ¼ ha1; fOrderg; fT itle; Userg; fBookg; fhto; grigi;
Intuitively, for each concept B 2 SB , there should be at
A2 ¼ ha2; fOrderg; fT itle; Userg; fMagazineg; fhto; grigi;
least one concept A 2 SA relevant to B, with respect to the
A3 ¼ ha3; fOrderg; fT itle; Userg; fMagazineg; fhto; ukigi;
filter set F . Otherwise, DLHset returns 0 (absolute mis-
A4 ¼ ha4; fSearchg; fT itleg; fBookg; fgi: match). The overall DLHset similarity is computed as the
We approach the SWS discovery problem as the mean value of the sum of the maximum DLHs for each
procedure of determining the similarity of an AðhI Da ; Ca ; concept B, since each B may have more than one relevant
I a ; Oa ; N F a iÞ and QðhI Dq ; Cq ; I q ; Oq ; N F q iÞ specification, concepts in SA .
based on three levels of similarity as follows: The Taxonomical Similarity denotes the similarity of two
specifications in terms of their concept membership sets Ci ,
1. Taxonomical Similarity (TS). It is computed over the
Ca and Cq sets of an A and Q specification and and therefore, it is equal to their DLHset similarity.
denotes their similarity in terms of their taxonomical Definition 7. The taxonomical similarity between A and
categorization in a Profile subclass hierarchy. Q specifications is defined as the DLHset similarity of their
2. Functional Similarity (FS). It is computed over the Ca and Cq sets, that is,
input (I a and I q ) and output (Oa and Oq ) sets of an
A and Q specification (signature similarity). T SðA; Q; FT Þ ¼ DLHset ðCq ; Ca ; FT Þ; ð3Þ
3. Nonfunctional Similarity (NFS). It is computed over
the values of the common nonfunctional properties where FT is the set of the hierarchical relationships that we
of an A and Q specification. allow to exist among the concepts of the Ca and Cq sets.
Authorized licensd use limted to: Imperial Coleg Lond. Downlade on June 07,21 at 19:3826 UTC from IE Xplore. Restricon aply.
282 IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, VOL. 22, NO. 2, FEBRUARY 2010
2 2 4 4 1
þ ECðOrder; P rofileÞ ¼ þ ¼ ; ¼1 ¼ :
5 5 5 5 5
Authorized licensd use limted to: Imperial Coleg Lond. Downlade on June 07,21 at 19:3826 UTC from IE Xplore. Restricon aply.
MEDITSKOS AND BASSILIADES: STRUCTURAL AND ROLE-ORIENTED WEB SERVICE DISCOVERY WITH TAXONOMIES IN OWL-S 283
For the other two specifications A2 and A3 , we have that is allowed to have 1) more general inputs than the
T SðA2 ; Q; FT Þ ¼ T SðA3 ; Q; FT Þ ¼ 1, only if e 2 FT , since query and 2) more general outputs than the query.
both specifications have the same taxonomical concepts to . Wsi ! FI ¼ FO ¼ fe; p; su; sig. This is the most re-
the Q specification. laxed filter, allowing also the existence of sibling
Note that both the EC and UC distances result in the relationships among I/O concepts.
same TS values for the example, since they are applied to The order of Web service filter relaxation is We < Wp <
the simple ontology of Table 1. In the general case, the Wsu < Wsi . Moreover, for each Wf filter, we define three
effectiveness of each measure depends on the characteristics additional levels of granularity based on the number of the
of each ontology. I/Os of the query and an advertisement. More specifically,
we define the exclusive (x), exclusive-input (xi), and exclusive-
4.2 The DLR Metric output (xo) grouping filters. The x filter is satisfied for the
The DLR metric denotes the similarity between A and Q matched advertisements that have the same number of I/
specifications in terms of the values in their common O parameters to the query. The xi filter is satisfied for the
properties. It is defined in terms of the Functional (FS) and advertisements that have the same number of input
Nonfunctional (NFS) similarities and a Web service filter Wf parameters only to the query. Similarly, the xo filter deals
that we analyze in the following sections: with the number of output parameters. For example, the
xWe filter matches advertisements that pass the We Web
Definition 8. Let two specifications A and Q. Their DLR service filter and have the same number of I/O parameters
similarity is the pair hF S; NF Si of their Functional and to the query. The xiWsi filter matches advertisements that
Nonfunctional similarities, with respect to the Web service pass the Wsi Web service filter and have the same number
filter Wf , that is, of input only parameters to the query input parameters.
The grouping filtering is motivated by the fact that an
DLRðA; Q; Wf Þ ¼ hF SðA; Q; Wf Þ; NF SðA; QÞi:
advertisement that satisfies, for example, the We filter and has
the same number of I/Os to the query, should be considered
4.2.1 Functional Similarity as a more “exact” result than an advertisement with different
number of I/O parameters. Based on this assumption, the
The FS is based on the DLHset similarity of the I/O sets of
matched advertisements are returned in groups (see Algo-
two specifications, so as to ensure that 1) all the advertise-
rithm 1 in Section 4.3), according to the Wf and grouping filter
ment inputs are satisfied by the query inputs and 2) all the
that satisfy. For each Wf , the order of relaxation
query outputs are satisfied by the advertisement outputs
is xWf < xiWf < xoWf < Wf . Therefore, xWe < xiWe <
(signature matching).
xoWe < We < xWp < xiWp < < xoWsi < Wsi (16 group-
Definition 9. The Functional Similarity between A and Q ing filters in total). The decision to define xiWf < xoWf is
specifications is the normalized value to ½0::1 that is defined, arbitrary.
with respect to the Web service filter Wf , as
Example. We exemplify on the computation of the FS of our
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi example, using the EC distance. Both Q and
F SðA; Q; Wf Þ ¼ DLHset ðI q ; I a ; FI Þ DLHset ðOa ; Oq ; Fo Þ:
A2 specifications have the same input sets I q ¼ I a ¼
ð5Þ fT itle; USERg w i t h DLHset ðI q ; I a ; FI Þ ¼ 1, only if
e 2 FI . For the output sets, we have that Oq ¼ fBookg
We use the geometric mean, instead of the arithmetic and Oa ¼ fMagazineg with
mean, because a Web service should be excluded if either of
ð2Þ
its input or output similarity is zero. DLHset ðOa ; Oq ; FO Þ ¼ DLHðBook; Magazine; FO Þ
In order to control the different degrees of relaxation ð1Þ
during I/O matching, the FS makes use of a Web service ¼ 1 ½ECðBook; ItemÞ
si2FO
filter Wf that defines the values of the hierarchical filter sets
þ ECðMagazine; ItemÞ
FI and FO in (5). More specifically, we define the Exact (We ),
Plugin (Wp ), Subsume (Wsu ), and Sibling (Wsi ) Web service 1 1 3
¼1 þ ¼ ;
filters with the following relationships to the FI and 5 5 5
FO filter sets:
only if si 2 FO , since the two concepts have a sibling
. We ! FI ¼ FO ¼ feg. This is the strictest filter that relationship and each concept is a direct subclass ðe ¼ 1Þ
allows two specifications to match only if they refer of the Item class. Therefore, from (5), we have that
qffiffiffiffiffiffiffiffi
to the same or to equivalent concepts in their I/Os. F SðA2 ; Q; Wf Þ ¼ 1 35 ¼ 0:774, only if we select the
. Wp ! FI ¼ fe; pg ^ FO ¼ fe; sug. This is a more re-
Wsi Web service filter that satisfies the hierarchical filter
laxed filter and intuitively denotes an A specification
that could be used instead of a Q specification. The requirements we have mentioned, that is, e 2 FI and
rationale is that all the inputs of the advertisement si 2 FO . Otherwise, F SðA2 ; Q; Wf Þ ¼ 0. The same holds
should be equivalent or subclasses of the query inputs, for the A3 specification. Furthermore, the A2 and
and all the outputs of the query should be equivalent A3 specifications satisfy the xWsi grouping filter, since
or superclasses of the advertisement outputs.
they both have the same number of input and output
. Wsu ! FI ¼ FO ¼ fe; p; sug. This filter relaxes even
more the matching criterion and the advertisement parameters to the Q specification.
Authorized licensd use limted to: Imperial Coleg Lond. Downlade on June 07,21 at 19:3826 UTC from IE Xplore. Restricon aply.
284 IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, VOL. 22, NO. 2, FEBRUARY 2010
The A4 specification has the {Title} input set, and For the A2 specification, we have that NF SðA2 ; QÞ ¼ 1,
therefore, its input DLH similarity to Q is DLHset ðfT itle; since they have the same value set for the object property to,
Userg; fT itleg; FI Þ ¼ 1, only if e 2 FI . Furthermore, both Q that is, a1:to ¼ a2:to ¼ fgrg. Finally, for the A4 specification,
and A4 have the same output sets ({Book}) and their output we have that a1:to ¼ fgrg 6¼ a4:to ¼ . In this way,
DLH similarity is DLHset ðfBookg; fBookg; FO Þ ¼ 1, only if NF SðA4 ; QÞ ¼ 0. Note that if we use the A4 specification
ep2 FO ffi . Therefore, from (5), we have that F SðA4 ; Q; Wf Þ ¼
ffiffiffiffiffiffiffiffi as a query, then all the NFS similarities would be equal to 1,
1 1 ¼ 1 that holds for each Web service filter Wf , since since A4 does not define any nonfunctional property, and
8Wf : e 2 FI ^ e 2 FO . Moreover, the A4 specification satis- therefore, Td [ To ¼ for each A specification in (6).
fies the xoWe grouping filter, since it has only the same
number of output parameters to the Q specification. 4.3 Overall Specification Similarity
The overall similarity sim of A and Q specifications is
4.2.2 Nonfunctional Similarity defined in terms of their TS, FS, and NFS similarities.
The NFS is defined in terms of two functions: the Definition 11. Let two specifications A and Q. Their similarity
dt function for computing the similarity of two data type sim is the triple hT S; F S; NF Si of their Taxonomical,
values and the ob function for computing the similarity of Functional, and Nonfunctional similarities, that is,
two object values, where dtða; bÞ and obða; bÞ 2 ½0::1. The
overall NFS similarity of two specifications is the mean simðA; Q; FT ; Wf Þ ¼
value of the dt and ob functions over the common properties hT SðA; Q; FT Þ; F SðA; Q; Wf Þ; NF SðA; QÞi:
of the two specifications.
Definition 10. Let the sets Td and To of the common data type The aggregation of the triple similarity into a single value
and object properties, respectively, of A and Q specifications is computed as the weighted mean sim of the three
with Td [ To 6¼ . The Nonfunctional Similarity is the similarities according to user requirements, that is,
normalized value to ½0::1 that is defined, with respect to the a T S þ b F S þ c NF S
functions dt and ob, as sim ¼ ;
aþbþc
NF SðA; QÞ ¼ where a, b, and c are normalized weights in ½0::1. The
P
8d2Td ½dt IDa :d; I Dq :d þ ob I Da :o; I Dq :o ð6Þ overall matchmaking algorithm of a Q specification with a
8o2To
: set of A specifications is depicted in Algorithm 1. The
jTd [ To j algorithm examines the complete set of the advertisements,
applying a two-phase filtering based on the taxonomical
If Td [ To ¼ , then we define that NF SðA; QÞ ¼ 1. The
and functional requirements. The rationale is to prune first
Td and To sets ignore the properties that have a value in an
the advertisements that do not taxonomically match with
A but not in a Q specification, assuming that requesters are
the query, in order for the more costly functional similarity
not interested in properties that do not annotate in queries.
procedure to be applied on a smaller set of advertisements.
NFS functions. The dt function computes the similarity
of two data type value sets VA and VQ by comparing Algorithm 1. The specification matchmaking algorithm
straightforwardly the values. In the case of xsd: string Require: a query advertisement Q, the set SA of the
ranges, we use the Jaro-Winkler similarity measure [17] advertised services, the taxonomical filter set FT , the
(str function), normalized to ½0::1 metric that calculates Web service filter Wf , the weights a; b; c 2 ½0::1 and two
the similarity of two strings as threshold filters lt ; lf 2 ½0::1
Ensure: an array of sets that contains the matched services
dtðVA ; VQ Þ ¼ max½strðva ; vq Þ;
8va 2VA
8vq 2VQ
grouped in each of the 16 grouping filters
1: matches
obtaining the most similar matching combination. In any 2: for all Ai 2 SA do
other case, we compare directly the two value sets, 3: ts T SðAi ; Q; FT )
returning a value between 0 and 1 that denotes the 4: If ts ¼ 0 _ ts < lt then
similarity of the sets in terms of the ratio of their common 5: continue
values, that is, 6: else
jVA \ VQ j 7: fs F SðAi ; Q; Wf )
dtðVA ; VQ Þ ¼ : ð7Þ 8: If fs ¼ 0 _ fs < lf then
jVA [ VQ j
9: continue
The ob function computes the similarity of two sets VA 10: else
and VQ of instances in the same way to (7). 11: wf getWf ðAi ; QÞ
Example. For the nonfunctional object property to of 12: nfs NF SðAi ; Q)
Table 1, the Q specification has the a1:to ¼ fgrg value set 13: end if
and the A3 specification the a3:to ¼ fukg value set. There- 14: end if
fore, we have that To ¼ ftog, Td ¼ and a ts þ b fs þ c nfs
15: sim
aþbþc
obðfukg; fgrgÞ jfukg \ fgrgj
NF SðA3 ; QÞ ¼ ¼ ¼ 0: 16: matches matches [ fhAi ; sim; wf ig
1 jfukg [ fgrgj
17: end for
Authorized licensd use limted to: Imperial Coleg Lond. Downlade on June 07,21 at 19:3826 UTC from IE Xplore. Restricon aply.
MEDITSKOS AND BASSILIADES: STRUCTURAL AND ROLE-ORIENTED WEB SERVICE DISCOVERY WITH TAXONOMIES IN OWL-S 285
18: G½16 ½; ; . An array of 16 sets that xoWe < xWsi (see Section 4.2.1). In other words, the
19: for all t hAk ; sim; wf i 2 matches do sim values are used to order the triples of the same
20: y wf getGroupingF ilterðQ; Ak Þ grouping set G½i, whereas the total ordering of the matches
21: G½y wf G½y wf [ ftg is defined based on the relaxation of the 16 grouping filters.
22: end for We argue that such an approach results in a better Top-k
23: for i 1; 16 do precision (Section 5.2), than that of following a total
24: G½i binarySortsim;desc ðG½iÞ ordering based on the sim values.
25: end for
26: Return G
5 EXPERIMENTAL RESULTS
More specifically, for each advertisement Ai of the SA set
(line 2), the algorithm computes first the TS (line 3). If the TS We tested OWLS-SLR and compared it to the OWLS-MX
equals 0 or it is less than the threshold lt (line 4), then Ai is [8] matchmaker (v1.1c), using the OWLS-TC version 2.2
ignored and the algorithm continues with the next revision 2 collection [18] with 1,007 OWL-S advertisements
and 29 queries. We have chosen OWLS-MX since it is a
specification. The threshold lt defines the minimum
well-known matchmaker, having been extensively tested
acceptable similarity between the taxonomical concepts,
on the OWLS-TC collection. Furthermore, it uses light-
allowing to incorporate different degrees of relaxation. The weight OWL-S SP descriptions like OWLS-SLR, it is able to
algorithm continues by computing the FS (line 7) using the incorporate the structural information of the domain
Web service filter Wf that is given as input to the algorithm. ontologies through concept unfolding (nearest neigh-
Similar to the TS, if the functional similarity equals 0 or it is bors—NN), and it is defined upon Pellet, the same reasoner
less than the threshold lf (line 8), the algorithm continues we also use. We used the M4 configuration of OWLS-MX
with the next specification. The lf threshold has similar role as the best configuration according to Klusch et al. [8]. The
to the lt threshold and defines the minimum acceptable OWLS-SLR configuration involved FT ¼ feg; a ¼ 0:8; b ¼ 1;
functional similarity. If the computed FS value is acceptable, c ¼ 0:1, and lt ¼ lf ¼ 0:5. We have chosen this configura-
then the algorithm retrieves the Web service filter wf that tion as the best one, after a number of experiments on the
the Q and A specifications satisfy (line 11). The algorithm collection. The experiments ran on a Windows XP PC with
continues with the computation of the NFS (line 12), and 3.2 GHz processor, setting maximum JAVA heap size of
finally, the sim value is computed (line 15) and the result is 800 Mbytes.
added to the set matches of the matched specifications as a
5.1 Loading and Query Response Time
triple of the matched specification, the sim value and the
The loading time involves the time needed to parse and
Web service filter wf (line 16).
process the advertisements and queries, whereas the query
The matched advertisements are returned according to
time involves the time needed to apply the matchmaking
the grouping filter that satisfies (line 20). More specifically,
algorithm. OWLS-SLR depicts a considerably better loading
each triple of the matches set is added to the G array
and query response performance compared to OWLS-M4.
(lines 20 and 21) that contains 16 sets, one set for each of the
OWLS-SLR loaded the data set in about 30 seconds,
16 grouping filters we have described in Section 4.2.1 (xWe ,
whereas OWLS-M4 needed more than 30 minutes. Fig. 2
xiWe , etc.). Finally, each set is ordered by the sim value of
depicts the query response times of OWLS-SLR and OWLS-
the triples (line 24) and G is returned.
M4. In OWLS-SLR, the UC-related distance is computed
Example. We exemplify on Algorithm 1, using the EC faster than the EC, since the latter requires the traverse of all
distance, setting a ¼ 0:8, b ¼ 1, and c ¼ 0:1, and ignoring the paths between the concepts. However, both configura-
the lt and lf parameters for simplicity. If we select FT ¼ tions perform faster than OWLS-M4 that depicts a constant
feg and Wf ¼ We , then none of the advertisements will be query response performance.
matched, since the A2 and A3 specifications will be pruned
si 5.2 Precision and Recall
during the computation of the FS (Book Magazine) and
the A4 specification will be pruned during the computa- We used the relevance sets of the collection in order to
si perform precision and recall tests. Due to the fact that the
tion of the TS (Search Order). By setting Wf ¼ Wsi , we
collection defines only direct Profile instances, we created
are able to retrieve the A2 and A3 specifications with also a taxonomy-based collection in order to test the
simA2 ¼ a1þb0:774þc1
aþbþc ¼ 0:881 and simA3 ¼ a1þb0:774þc0
aþbþc ¼ performance taking into account the TS. Note that OWLS-
0:828 (see the examples in Sections 4.1.1, 4.2.1, and 4.2.2 MX can handle only direct Profile instances. Fig. 3 depicts
for the computation of the similarity values). Furthermore, the average precision and Fig. 4 the average recall of all
both specifications satisfy the xWsi grouping filter (same queries for OWLS-M4 and OWLS-SLR, according to the
number of I/Os to Q), and therefore, G½xWsi ¼ Web service filter that was used. We have omitted the
fhA2 ; 0:881; xWsi i; hA3 ; 0:828; xWsi ig. subsumed-by filter of OWLS-MX for presentation purpose.
OWLS-M4 has, in general, better precision than OWLS-
Moreover, by setting FT ¼ fe; sig, we are able to retrieve
a1þb1þc0 SLR in the collection without a Profile taxonomy, showing
also the A4 specification with simA4 ¼ 5aþbþc ¼ 0:61 that the filter definitions that it follows, which are based
and G½xoWe ¼ fhA4 ; 0:61; xoWe ig. Note that, even if the on [12], fit better to the specific collection. However, by
sim value of A4 is less than the sim values of the other two performing domain-oriented discovery, OWLS-SLR out-
specifications, A4 will be returned as a more relevant match, performs OWLS-M4, justifying the advantage of a do-
followed by A2 and A3 , since the order of relaxation defines main-oriented approach to SWS discovery. The recall of
Authorized licensd use limted to: Imperial Coleg Lond. Downlade on June 07,21 at 19:3826 UTC from IE Xplore. Restricon aply.
286 IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, VOL. 22, NO. 2, FEBRUARY 2010
OWLS-SLR using the Profile taxonomy is a little bit lower Fig. 6 depicts the scalability of OWLS-SLR in terms of loading
than the one without taxonomy, since some results do not and Fig. 7 depicts the scalability in terms of query response
pass the FT ¼ feg filter set. time. For the legibility of the presentation, Fig. 7 depicts only
In most cases, we are interested in the first k results of a the query response times up to 4,028 advertisements, using
query. Fig. 5 depicts the average precision of all queries for the UC-related distance. OWLS-SLR scales very well (almost
OWLS-SLR and OWLS-M4 at the Top-k places. The linearly) both on loading and query response times.
precision at Top-k for a query q is computed as
5.2.2 Discussion
Relevantq \ Returnedq;k
; We believe that the approach of OWLS-MX to maintain and
Returnedq;k modify a local ontology imposes an extra overhead on the
where Relevantq is the relevance set of q and Returnedq;k is loading performance. OWLS-SLR loads directly the Profile
the Top-k results of the returned advertisements. The instances as well as the domain ontologies into the reasoner,
experiments have shown that OWLS-SLR has a consider- and therefore, any loading overhead is only related to the
capabilities of the underlying reasoner. Furthermore,
ably better precision than OWLS-M4 on the results that are
OWLS-MX performs a concept unfolding for determining
returned first.
concept similarities in the case of nearest neighbor (sibling)
5.2.1 Scalability matches, generating vectors on which the IR techniques are
applied. The determination of the sibling relationships in
In order to test the scalability of OWLS-SLR, we generated OWLS-SLR is performed directly on the reasoner’s graph
synthetic data sets by altering the base URI of advertisements. that seems to be a more efficient and scalable approach.
Regarding precision and recall, the strong point of
OWLS-SLR is that it allows the existence of a Profile
Authorized licensd use limted to: Imperial Coleg Lond. Downlade on June 07,21 at 19:3826 UTC from IE Xplore. Restricon aply.
MEDITSKOS AND BASSILIADES: STRUCTURAL AND ROLE-ORIENTED WEB SERVICE DISCOVERY WITH TAXONOMIES IN OWL-S 287
taxonomy in contrast to OWLS-MX that handles only direct where Cm are domain concepts. Practically, an RAC is an
Profile instances. To demonstrate the effectiveness of the TS ontology concept that is defined using roles and concepts.
in matchmaking, we present an example taken from OWLS- In fact, our RAC-based approach tries to incorporate the CC
TC. EBookOrder1 is a query of the collection with the object modeling capabilities into the SP paradigm. However,
specification hq; fEconomyg; fT itle; Userg; fBookg; fgi. instead of considering a Web service as a whole, we give
Without taking into account the Profile taxonomy, the the opportunity to treat a particular input or output
EBookOrder1 query matches the BookFinder advertisement, parameter as a whole.
which is defined as hadv; fEducationg; fT itleg; fBookg; fgi,
in both OWLS-SLR and OWLS-MX. However, these two 6.2 Example and Experiments
specifications belong to different domains (Economy and The RAC-oriented discovery requires the runtime classifi-
Education) and the relevance set of EBookOrder1 does not cation of RACs in domain ontologies. To this end, we
contain the BookFinder advertisement. By considering the enhanced OWLS-SLR with the ability of altering at runtime
Profile taxonomy in OWLS-SLR, the BookFinder advertise- the domain ontologies. However, all the query RACs are
ment is not returned in our experiments, since FT ¼ feg. removed from the domain ontologies after the application
Furthermore, OWLS-SLR orders the results based on of the matchmaking algorithm, preserving the initial
grouping filters (Section 4.2.1). OWLS-MX does not perform ontology structure.
grouping, returning the results based on a total similarity We extended the OWL-S Process concept so as to allow
ordering. In this way, OWLS-SLR has better Top-k precision OWLS-SLR to handle Web service Profile instances that
since some matches are considered more relevant to some contain either RACs or ordinary ontology concepts. More
others, even if they satisfy the same Web service filter. specifically, we defined the rac: Input and rac: Output
concepts as specializations to the process:Input and
6 ROLE-DRIVEN WEB SERVICE DISCOVERY process: Output concepts. We defined also the roles
rac:concept and rac: minCardinalityRestriction in order to
The OWL-S SP paradigm uses predefined ontology concepts enable the definition of Cm s and role restrictions in (10).
to annotate the Web service I/O parameters. However, it is We exemplify on the way ontology roles can be used as
impractical and not realistic to assume that there will always annotation constructs, using the (input) RAC-oriented
be an ontology concept suitable for our needs. The ontology advertisements of Table 2 that describe Web services that
roles are also important modeling constructs that encapsu- return the price of books. For simplicity, all the outputs
late domain knowledge. Bear in mind that the CC approach have been annotated with the P rice concept. In A1 , the
makes extensive use of ontology roles. We exploit the input is annotated using the equivalent RAC to the Book
classification capabilities of DL reasoning and enhance our
framework with the ability to perform SWS discovery using
also ontology roles as annotation constructs. TABLE 2
RAC-Oriented Specifications
6.1 Role-Oriented Annotation Concepts
We introduce the notion of the Role-oriented Annotation
Concept (RAC), a specially defined concept that is derived
from cardinality restrictions on ontology roles.
Definition 12. Let R be a set of ontology roles. The Role-oriented
Annotation Concept for the set R is the equivalent concept to
the intersection of minimum cardinality restrictions of the form
Authorized licensd use limted to: Imperial Coleg Lond. Downlade on June 07,21 at 19:3826 UTC from IE Xplore. Restricon aply.
288 IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, VOL. 22, NO. 2, FEBRUARY 2010
Fig. 8. The RAC-based against the typical query response time of OWLS-SLR and OWLS-MX.
concept. This is similar to the SP paradigm, using directly allowing both concepts and roles to be considered in the
an ontology concept. In A2 , the input is annotated with an annotation process. It is worth mentioning that the RAC-
RAC based on the role title, and A3 defines an RAC using based query answering performance of OWLS-SLR is better
the role isbn. Finally, A4 defines an input RAC using both than that of OWLS-MX.
the title and isbn roles.
By classifying the RACs into the domain ontology of
Table 2, we obtain the TBox relationships of Fig. 9.
7 RELATED WORK
Let the Q specification of Table 2. By classifying RAC E in OWL-S [5], SAWSDL [19], WSDL-S [20], and WSMO [9]
the domain ontology, we get that RAC E RAC B . Therefore, constitute the major standards for semantic Web service
if we use the We filter during matchmaking, only the annotation. Apart from the Light approach we described in
A2 specification will be returned, since it matches exactly all Section 2.1.1, WSMO-DF allows the definition of descrip-
the I/Os of Q. By relaxing the Web service filter, we retrieve tions with higher level of detail. The WSMO capability
the other three A specifications as well. More specifically, the element is able to encapsulate goals, mediators, precondi-
p
Wp filter returns A4 , since RAC D RAC E ; the Wsu filter tions, and assumptions. It is argued that a similar level of
su
returns A1 , since RAC A RAC E ; and the Wsi filter returns detail can also be used in OWL-S [21], [22] through, for
si
A3 , since RAC A RAC E . In this way, ontology roles can be example, the Process model [23] or preconditions and
used as annotation concepts in matchmaking. effects [24], [25]. A lightweight approach, such as ours, can
In order to test the performance of the RAC-based SWS be used as an initial step to retrieve a set of candidate Web
matchmaking, we used the OWLS-TC collection, and for services on where more complex and sophisticated algo-
each I/O query concept C, we defined the equivalent RAC, rithms can be applied.
that is, RAC C, in a similar way to the A1 RAC-oriented In [26], Web services are annotated in terms of state
advertisement of Table 2. In this way, we were able to use transitions. Actually, the Rich Web service representation of
the same relevance set to the initial collection, as well as to WSMO-DF is followed, using an environmental ontology.
test the performance of the runtime concept classification. In contrast to [26], we follow the SP model, exploiting the
As far as precision and recall are concerned, the structural ontology information.
experiments resulted in the same performance that we IRS-III [27] extends the WSMO conceptual model and
achieved in OWLS-SLR using the typical query collection uses the OCML language for internal representation and an
(Figs. 3 and 4). This fact justifies the soundness of the RAC- OCML reasoner. In contrast to our framework, IRS-III
based implementation of OWLS-SLR. follows the Rich WSMO model and a frame-based rule
Regarding query response time, OWLS-SLR requires language for representing ontologies.
more time to answer the RAC-based queries than in the A search engine for grid service discovery is presented in
typical collection, as it is depicted in Fig. 8. This happens [28], using the Rough sets theory. The novelty is in its
since the computation of the FS similarity involves also the capability to deal with uncertain properties, that is, proper-
time Pellet requires to classify and delete the RAC concepts, ties that are explicitly used by one advertisement but do not
in contrast to the typical OWLS-SLR functionality, where appear in another service of the same category. In our
the TBox reasoning is performed in advance. However, we framework, we examine only the common properties of an
believe that the role-driven SWS discovery offers more advertisement and query.
capabilities than the typical OWL-S SP-oriented paradigm, In [29], an approach based on context-aware ratings and
context-aware experiences is proposed in order to select
services. Consumers use ontologies to express the context of
their interactions with service providers as a whole, instead
of using the SP. It targets mainly at rating environments
without exploiting structural knowledge, but it may be used
in Web service discovery.
In [30], Web service descriptions are defined as CCs in
OWL and the matchmaking procedure examines the
subsumption relationships. FC-MATCH [31] follows the
Fig. 9. The TBox relationships among the RACs. same approach, performing also text similarity matching
Authorized licensd use limted to: Imperial Coleg Lond. Downlade on June 07,21 at 19:3826 UTC from IE Xplore. Restricon aply.
MEDITSKOS AND BASSILIADES: STRUCTURAL AND ROLE-ORIENTED WEB SERVICE DISCOVERY WITH TAXONOMIES IN OWL-S 289
using WordNet. In [32], a framework for annotating Web applied on a smaller set of candidate Web service descrip-
services using DLs is presented. Similar to ours, it follows tions than the complete initial set.
the abstract Web service model. However, it treats Web In an effort to enhance the instance-oriented SP dis-
services as a whole. covery paradigm with the domain modeling capabilities
In the DAML-S/UDDI Matchmaker [33], OWL-S SP found on CC approaches, 1) we allow the existence of
advertisements and requests refer to DAML concepts and Profile taxonomies and 2) we enable the annotation of I/Os
the matching process performs inferences on the subsump- with ontology roles. Moreover, we defined a matchmaking
tion hierarchy. It uses a different definition of Web service algorithm that exploits the structural knowledge of ontol-
filters from ours and it does not consider Profile taxo- ogies, for example, sibling concept relationships, by con-
nomies, roles, or grouping filtering. sidering advertisements and requests as objects and
LARKS [34] uses both syntactic and semantic match- implementing concept (dis-) similarity measures.
ings. It uses five matchmaking filters, namely context We presented also a comparison of OWLS-SLR to OWLS-
matching, profile comparison, similarity matching, signa- MX. The experiments have shown a considerably better
ture matching, and constant matching. LARKS uses its performance on loading and querying of OWLS-SLR than
own capability description and DL language in contrast to that of OWLS-MX. Furthermore, we were able to increase
our OWL-based approach. precision and recall using grouping filters (Top-k experi-
OWLS-MX [8] utilizes both logic-based reasoning and ments) and performing taxonomy-based discovery. The
content-based IR techniques for Web services in OWL-S. As results seem very promising as far as the requirement for a
we have already mentioned, it cannot handle Profile fast filtering of Web service advertisements is concerned.
taxonomies and it follows the static SP paradigm, being OWLS-SLR is available in [42], together with the
unable to use dynamically ontology roles. iMatcher2 [35] experimental collections we have used. For the future, we
follows the OWLS-MX approach, applying also learning plan to enhance our framework with more structural [43]
algorithms in order to predict similarities. Like OWLS-MX, and information-content similarity measures [44], [45], [46].
it uses a DL reasoner in order to unfold the annotation Currently, we are working on enhancing our framework
concepts, creating a vector on which the IR techniques are with composition capabilities in order to return not only
applied. iMatcher2 does not follow a standard matchmak- single Web services, but also Web service compositions [47]
ing algorithm, which is defined through an iSPARQL based on abstract descriptions.
strategy. WSMO-MX [36] is a hybrid approach based on
Rich WSMO service descriptions. ACKNOWLEDGMENTS
There are plenty of other approaches that are based on I/
This work was partially supported by a PENED program
Os, for example, [37], [38], [25]. To the best of our
(EPAN M.8.3.1, No. 0373), jointly funded by the EU and
knowledge, these approaches do not perform ABox reason-
the General Secretariat of Research and Technology.
ing on Profile instances in order to exploit the domain
knowledge of Profile instances. Instead, they retrieve
directly the I/O annotations and any taxonomical knowl- REFERENCES
edge stems from special properties, such as service [1] WSDL 1.1, http://www.w3.org/TR/wsdl, 2001.
categorization. Furthermore, they do not consider roles, [2] M. Burstein, C. Bussler, M. Zaremba, T. Finin, M.N. Huhns, M.
Paolucci, A.P. Sheth, and S. Williams, “A Semantic Web Services
using static annotation concepts, and do not apply further Architecture,” IEEE Internet Computing, vol. 9, no. 5, pp. 72-81,
filtering on results (grouping filtering). METEOR-S [39] Sept./Oct. 2005.
follows the WSDL-S approach, where WSDL constructs [3] C. Preist, “A Conceptual Architecture for Semantic Web Services,”
Proc. Int’l Semantic Web Conf., pp. 395-409, 2004.
point to ontology concepts.
[4] U. Keller, R. Lara, H. Lausen, and D. Fensel, Semantic Web Service
Our work has been motivated by a previous work of ours Discovery in the WSMO Framework. Idea Publishing Group, 2006.
[40] that implements Object-Oriented similarity measures [5] D. Martin, M. Burstein, D. Mcdermott, S. Mcilraith, M. Paolucci, K.
for SWS discovery, using a production rule engine [41]. In Sycara, D.L. Mcguinness, E. Sirin, and N. Srinivasan, “Bringing
Semantics to Web Services with OWL-S,” World Wide Web, vol. 10,
the present work, 1) we use a DL reasoner to handle SWS no. 3, pp. 243-277, Sept. 2007.
descriptions and apply an extended matchmaking algo- [6] R. Bergmann and M. Schaaf, “Structural Case-Based Reasoning
rithm, 2) we define two new OWL-S Profile-aware and DL- and Ontology-Based Knowledge Management: A Perfect Match?”,
J. Universal Computer Science (UCS), vol. 9, no. 7, pp. 608-626, 2003.
based similarity measures, and 3) we propose a framework [7] F. Baader, The Description Logic Handbook: Theory, Implementation
for the incorporation of ontology roles in the SP SWS and Applications. Cambridge Univ. Press, Jan. 2003.
discovery paradigm. [8] M. Klusch, B. Fries, and K. Sycara, “OWLS-MX: A Hybrid
Semantic Web Service Matchmaker for OWL-S Services,” Web
Semantics: Science, Services, and Agents on the World Wide Web,
8 CONCLUSIONS AND FUTURE WORK vol. 7, no. 2, pp. 121-133, Apr. 2009.
[9] Enabling Semantic Web Services: The Web Service Modeling Ontology,
In this paper, we presented an OWL-S SP-aware framework D. Fensel, H. Lausen, A. Polleres, J.D. Bruijn, M. Stollberg,
D. Roman, and J. Domingue, eds. Springer-Verlag, 2006.
for SWS discovery, using abstract and lightweight Web
[10] OWL, http://www.w3.org/2004/OWL/, 2004.
service descriptions. Our intention is to define a framework [11] OWL-S 1.1 Release: Examples, http://www.daml.org/services/
that can be used as a prephase in more fine-grained owl-s/1.1/examples.html, 2004.
approaches that incorporate rich Web service descriptions, [12] A.M. Zaremski and J.M. Wing, “Specification Matching of Soft-
ware Components,” Proc. Third ACM SIGSOFT Symp. Foundations
such as preconditions, effects, or state transitions. In this of Software Eng., pp. 6-17, 1995.
way, the complex and sophisticated algorithms would be [13] UDDI, www.oasis-open.org/committees/uddi-spec, 2005.
Authorized licensd use limted to: Imperial Coleg Lond. Downlade on June 07,21 at 19:3826 UTC from IE Xplore. Restricon aply.
290 IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, VOL. 22, NO. 2, FEBRUARY 2010
[14] N. Srinivasan, M. Paolucci, and K.P. Sycara, “An Efficient [39] K. Verma, K. Sivashanmugam, A. Sheth, A. Patil, S. Oundhakar,
Algorithm for OWL-S Based Semantic Search in UDDI,” Semantic and J. Miller, “METEOR-S WSDI: A Scalable P2P Infrastructure of
Web Services and Web Process Composition, pp. 96-110, Springer, Registries for Semantic Publication and Discovery of Web
2004. Services,” J. Information Technology and Management, vol. 6, no. 1,
[15] A. Maedche and V. Zacharias, “Clustering Ontology-Based pp. 17-39, 2005.
Metadata in the Semantic Web,” Proc. European Conf. Principles of [40] G. Meditskos and N. Bassiliades, “Object-Oriented Similarity
Data Mining and Knowledge Discovery, pp. 348-360, 2002. Measures for Semantic Web Service Matchmaking,” Proc. European
[16] E. Sirin, B. Parsia, B.C. Grau, A. Kalyanpur, and Y. Katz, “Pellet: A Conf. Web Services, pp. 57-66, 2007.
Practical OWL-DL Reasoner,” J. Web Semantics, vol. 5, no. 2, [41] G. Meditskos and N. Bassiliades, “A Rule-Based Object-Oriented
pp. 51-53, 2007. OWL Reasoner,” IEEE Trans. Knowledge and Data Eng., vol. 20,
[17] W. Winkler, “The State of Record Linkage and Current Research no. 3, pp. 397-410, Mar. 2008.
Problems,” Proc. Conf. Survey Methods Section, Statistical Soc. [42] OWLS-SLR, http://lpis.csd.auth.gr/systems/OWLS-SLR, 2008.
Canada, pp. 73-80, 1999. [43] V. Schickel-Zuber and B. Faltings, “OSS: A Semantic Similarity
[18] OWLS-TC Version 2.2 Revision 2, http://projects.semwebcentral. Function Based on Hierarchical Ontologies,” Proc. Int’l Joint Conf.
org/projects/owls-tc/, 2008. Artificial Intelligence, pp. 551-556, 2007.
[19] J. Kopeck y, T. Vitvar, C. Bournez, and J. Farrell, “SAWSDL: [44] P. Resnik, “Using Information Content to Evaluate Semantic
Semantic Annotations for WSDL and XML Schema,” IEEE Internet Similarity in a Taxonomy,” Proc. Int’l Joint Conf. Artificial
Computing, vol. 11, no. 6, pp. 60-67, Nov./Dec. 2007. Intelligence, pp. 448-453, 1995.
[20] WSDL-S, http://www.w3.org/Submission/WSDL-S/, 2005. [45] D. Lin, “An Information-Theoretic Definition of Similarity,” Proc.
[21] M. Paolucci, N. Srinivasan, and K. Sycara, “Expressing WSMO Int’l Conf. Machine Learning, pp. 296-304, 1998.
Mediators in OWL-S,” Proc. Third Int’l Semantic Web Conf., [46] P.M. Schwarz, Y. Deng, and J.E. Rice, “Finding Similar Objects
Semantic Web Services: Preparing to Meet the World of Business Using a Taxonomy: A Pragmatic Approach,” Proc. Int’l Conf.
Applications, 2004. Databases and Applications of Semantics, pp. 1039-1057, 2006.
[22] R. Lara, D. Roman, A. Polleres, and D. Fensel, “A Conceptual [47] U. Küster, B. König-Ries, M. Stern, and M. Klein, “DIANE: An
Comparison of WSMO and OWL-S,” Proc. European Conf. Web Integrated Approach to Automated Service Discovery, Match-
Services, pp. 254-269, 2004. making and Composition,” Proc. Int’l Conf. World Wide Web,
[23] M. Klein and A. Bernstein, “Toward High-Precision Service pp. 1033-1042, 2007.
Retrieval,” IEEE Internet Computing, vol. 8, no. 1, pp. 30-36, Jan.
2004. Georgios Meditskos received the PhD degree
[24] M. Klein and B. König-Ries, “Coupled Signature and Specification in computer science in 2009 from the Depart-
Matching for Automatic Service Binding,” Proc. European Conf. ment of Informatics, Aristotle University of
Web Services, pp. 183-197, 2004. Thessaloniki, Greece, for his dissertation on
[25] D. Skoutas, A. Simitsis, and T. Sellis, “A Ranking Mechanism for “Semantic Web Service Discovery and Ontology
Semantic Web Service Discovery,” Proc. IEEE Congress on Services, Reasoning using Entailment Rules.” He also
pp. 41-48, 2007. received the MSc and a BSc degrees from the
[26] P. Wang, Z. Jin, L. Liu, and G. Cai, “Building Toward Capability same department. He is interested in semantic
Specifications of Web Services Based on an Environment Web technologies and, more specifically, in the
Ontology,” IEEE Trans. Knowledge and Data Eng., vol. 20, no. 4, combination of ontologies and rules. His re-
pp. 547-561, Apr. 2008. search has also been focused on the semantic Web service discovery
[27] J. Domingue, L. Cabral, S. Galizia, V. Tanasescu, A. Gugliotta, B. and composition.
Norton, and C. Pedrinaci, “IRS III: A Broker-Based Approach to
Semantic Web Services,” J. Web Semantics, vol. 6, no. 2, pp. 109-132,
2008.
[28] M. Li, B. Yu, O.F. Rana, and Z. Wang, “Grid Service Discovery Nick Bassiliades received the PhD degree in
with Rough Sets,” IEEE Trans. Knowledge and Data Eng., vol. 20, parallel knowledge base systems from the
no. 6, pp. 851-862, June 2008. Department of Informatics at the Aristotle Uni-
[29] M. Şensoy and P. Yolum, “Ontology-Based Service Representation versity of Thessaloniki, Greece, in 1998. He is
and Selection,” IEEE Trans. Knowledge and Data Eng., vol. 19, no. 8, currently an assistant professor in the Depart-
pp. 1102-1115, Aug. 2007. ment of Informatics at the Aristotle University of
[30] L. Li and I. Horrocks, “A Software Framework for Matchmaking Thessaloniki. His research interests include
Based on Semantic Web Technology,” Proc. Int’l Conf. World Wide knowledge-based systems, agents, and the
Web, pp. 331-339, 2003. semantic Web. He has published more than
[31] D. Bianchini, V.D. Antonellis, M. Melchiori, and D. Salvi, 100 papers in journals, conferences, and books,
“Semantic-Enriched Service Discovery,” Proc. Int’l Conf. Data and coauthored an international book on parallel, object-oriented, and
Eng. Workshops, p. 38, 2006. active knowledge-based systems and a Greek book on artificial
[32] S. Grimm, B. Motik, and C. Preist, “Matching Semantic Service intelligence. He is a member of the Board of the Greek Artificial
Descriptions with Local Closed-World Reasoning,” Proc. European Intelligence Society, a director of RuleML, Inc., and also a member of
Semantic Web Conf., pp. 575-589, 2006. the Greek Computer Society, the IEEE, and the ACM.
[33] K.P. Sycara, M. Paolucci, A. Ankolekar, and N. Srinivasan,
“Automated Discovery, Interaction and Composition of Semantic
Web Services,” J. Web Semantics, vol. 1, no. 1, pp. 27-46, 2003. . For more information on this or any other computing topic,
[34] K. Sycara, S. Widoff, M. Klusch, and J. Lu, “LARKS: Dynamic please visit our Digital Library at www.computer.org/publications/dlib.
Matchmaking among Heterogeneous Software Agents in Cyber-
space,” Autonomous Agents and Multi-Agent Systems, vol. 5, no. 2,
pp. 173-203, 2002.
[35] C. Kiefer and A. Bernstein, “The Creation and Evaluation of
iSPARQL Strategies for Matchmaking,” Proc. European Semantic
Web Conf., pp. 463-477, 2008.
[36] F. Kaufer and M. Klusch, “WSMO-MX: A Logic Programming
Based Hybrid Service Matchmaker,” Proc. European Conf. Web
Services, pp. 161-170, 2006.
[37] J. Cardoso, “Discovering Semantic Web Services with and without
a Common Ontology Commitment,” Proc. IEEE Services Computing
Workshops, pp. 183-190, 2006.
[38] J. Pathak, N. Koul, D. Caragea, and V.G. Honavar, “A Framework
for Semantic Web Services Discovery,” Proc. ACM Int’l Workshop
Web Information and Data Management, pp. 45-50, 2005.
Authorized licensd use limted to: Imperial Coleg Lond. Downlade on June 07,21 at 19:3826 UTC from IE Xplore. Restricon aply.