WSRS: A WEB SERVICE RECOMMENDER SYSTEM
Esma Aïmeur, David Daboue, Flavien Serge Mani Onana
University of Montreal, Department of Computer Science, Montreal, Canada
Djamal Benslimane
Claude Bernard University, Lyon 1, Villeurbanne, France
Zakaria Maamar
Zayed University, Dubai, United Arab Emirates
Keywords: Profile, Recommender System, Similarity, Web service.
Abstract: Despite Web services widespread adoption, users still struggle with the problem of locating the Web
services that best satisfy their needs and meet their requirements. Unfortunately, current Web services
repositories suffer from various limitations, such as providing Web services to users regardless of these
users’ past experiences and these Web services’ intrinsic characteristics like popularity and credibility. In
this paper, we introduce a Web Service Recommender System that uses collaborative filtering, demographic
filtering, and content-based filtering techniques to help facilitate the search of Web services. The WSRS
system uses both users’ and Web services’ profile for the sake or recommending Web services.
1 INTRODUCTION
The W3C introduced Web services to provide a
standard way of communication between software
applications. Web services are technologies
designed to support interoperable computer-to-
computer interaction over networks.
Like a company wishing to advertise and make
its activities more visible to prospective users, the
need for publishing Web services for discovery
needs draws much attention from several research
bodies. To this end, UDDI enables applications to
publish and find Web services. Indeed, today, it is a
challenge for application designers and simple users
to find the Web services that best satisfy their needs
and meet their requirements. Usually, techniques to
search for Web services in UDDI repositories are
keyword-based. Tsalgatidou et al. (2002) argue that
even if the WSDL/UDDI contained semantic
information in connection with the providers and
syntactic information in connection with the Web
services, it remains difficult to decide how to find
“relevant” Web services, starting from simple
keywords. Therefore, there is a need to design
systems that assist users locate Web services. In
particular, such systems may take into account the
user profile, which summarizes what a user likes and
dislikes. This profile includes the user’s
demographic data, interests, preferences, usage
records, purchase records, browsing behavior, etc.
Usually, the profile is used for Web personalization,
which is an Internet technique for adapting websites
to individuals (Mobasher and Anand, 2005).
Nowadays, there exist several ways to compile user
profiles (Turban et al., 2006). However, the different
approaches should consider the “size and the
heterogeneous nature of the data itself, as well as the
dynamic nature of user interactions with the Web”.
In this paper, we compile the profile of users and
Web services in order to match Web services’
capabilities to users’ needs. Thus, we concentrate on
the approach based on recommender systems, which
make inferences from data provided by users on
other issues or by analyzing similar users.
Recommender systems allow entities providing
items to guide the choices made by users. The
recommendation can be issued by the service
provider who possesses the items or by brokers. The
284
Aïmeur E., Daboue D., Serge Mani Onana F., Benslimane D. and Maamar Z. (2007).
WSRS: A WEB SERVICE RECOMMENDER SYSTEM.
In Proceedings of the Third International Conference on Web Information Systems and Technologies - Internet Technology, pages 284-289
DOI: 10.5220/0001288102840289
Copyright
c
 SciTePress

recommender system we introduce, here, is a search
tool that can be anchored to any UDDI.
2 PROFILING APPROACH
2.1 How to Profile Users?
Usually, a Web user could be defined with three
types of information: demographic, usage behavior
(what the user bought in the past, for how much,
etc.), and browsing behavior (click stream). The
demographic information is provided by the user
when he opens an account in the system being used.
The user can then update this information during his
subsequent visits. For the last type of information, as
the user progressively browses the system’s website,
information about the items he is testing or using, is
collected and then stored for further use. This kind
of information is known as implicit feedback that the
system gathers from the user’s browsing behavior.
Moreover, the user is usually invited to provide a
rating on a scale for any items that he tests or uses.
This rating is an explicit feedback that the system
also stores for further use. Browsing behavior
information and ratings are data that the system
gathers on each user to enrich his profile and use
later during recommendation. The usage behavior
information is an implicit feedback gathered from
the user’s activities in the system. Each user profile
contains a list of Web services he used in the past.
Moreover, the user profile includes the usage
percentage, which expresses the number of times
that a given Web service has been accessed by the
user, session after session. It represents the fidelity
of the user in using a given Web service.
2.2 How to Profile Web services?
Besides its operations, input parameters and output,
what are other characteristics that may distinguish a
Web service from other peers? Moreover, this
question leads to the notion of Web service profile.
We suggest profiling a Web service with a set of
characteristics, which could to a certain extent
overlap with what is usually known as non-
functional aspects of Web services: Popularity
(number of times the Web service has been used
with success), Response delay, Intelligibility (how
comprehensible is a given Web service), Credibility,
Lifetime, Updatability (does the Web service take
into account users’ feedback?).
This paper only focuses on popularity and an
average rating that mixes the aforementioned
characteristics (response delay, intelligibility, etc.).
3 RECOMMENDER SYSTEMS
The main interest in recommender systems is backed
by the plethora of applications dealing with
information overload so personalized content and
services are provided to users A recommender
system relies on an item's features and/or previous
user ratings to provide an opinion or a list of
selected items that assists the user in evaluating
items that are not yet rated by him. Five main types
of filtering techniques are identified in (Burke,
2002). We define below the three most used
techniques, as we use them in our recommender
system: Collaborative Filtering, Content-based
Filtering, and Demographic Filtering.
3.1 Collaborative Filtering
Collaborative Filtering (CF) was first introduced by
Tapestry’s developers (Goldberg et al., 1992). It
accumulates user item ratings, identifies users with
common ratings and offers recommendations based
on inter-user comparison. CF techniques are
becoming increasingly important in the context of e-
commerce, with the unprecedented growth in the
number of users.
In this paper, we use a well-known memory-
based (as opposed to a model-based) (Breese et al.,
1998) CF technique, based on the Nearest Neighbor
classification algorithm. However, our general
approach to Web service recommendation is not tied
to this choice. Here, data is represented by a matrix
where entry v
u,i
represents the rating user u gave to
item (Web service) i. This entry is set to null in case
user u has not rated yet item i in which case this
entry is not used in the computation. Suppose that
the Web service provider's database, T, contains t
items, p
1
, p
2
,…, p
t
, and that m users, u
1
, u
2
,…, u
m
,
have rated some items from T. Rating predictions for
a given user are produced in two stages, as we now
review.
In order to estimate similarity between users, various
metrics have been proposed. One of them, for
example, is the Pearson correlation (Resnick et al.,
1994). The results obtained range from -1 for
negative correlation to +1 for perfect positive
correlation. Specifically, let equation (1) stand for
the correlation between users c and u, where J is the
set of items rated by both users c and u, v
l,j
is the
rating user l gives to item j and
l
v
is the average
rating of user l for the items that belong to J, for
WSRS: A WEB SERVICE RECOMMENDER SYSTEM
285

}{
ucl ,∈
.
∑∑
∑
∈∈
∈
−−
−−
=
Jj
u
ju
Jj
c
jc
Jj
u
ju
c
jc
vvvv
vvvv
uccorr
2
,
2
,
,,
)()(
))((
),(
(1)
The weighted sum equation below can then be
used to predict the rating of user c on item i. The
resulting predictions are sorted and those with
highest values are selected for recommendation.
∑
∑
∈
∈
−
+=
Uu
Uu
uju
cjc
uccorr
vvuccorr
vP
),(
))(,(
,
,
(2)
In equations (1) and (2), U is the set of all users who
rate item i. (In particular, user c for whom
predictions are being computed, does not belong to
U). These equations are used in the recommendation
process (Section 4.2).
3.2 Content-based Filtering
When Content-based Filtering (CN) techniques are
used, items are compared based on their content,
which can be described using explicit features. The
description can also consist of textual documents
with their titles, illustrations, tables of contents, etc.
In this paper, items are Web services consisting of
one or several operations, which we consider as
features. For its part, an operation consists of name,
input parameters, output, and description. Thus,
comparing two Web services is like finding the
similarity between their operations. A recommender
system using CN learns a user's interests from the
description of the items the user rates. This enables
the system to profile users (Pazzani, 1997). As was
the case for collaborative filtering, such profiles are
long-term user models. These models can be
updated as long as users rate items and implicitly
change their preferences. In order to estimate
similarity between items, various metrics have been
proposed (e.g. (Schafer et al., 1999)). In our case,
we use similar_text (string first, string second), a
well-known PHP function, originally proposed by
Oliver (Oliver, 1993), to compute the similarity
between Web services (Section 4.2).
3.3 Demographic Filtering
Recommender systems based on Demographic
Filtering (DF) aim at categorizing users based on
their demographic information and recommend
items accordingly. More precisely, demographic
information is used to identify the types of users that
like similar items. The key element of DF is that it
creates categories of users having similar
demographic characteristics, and tracks the
aggregate usage behavior or preferences of users
within these categories. Recommendations for a new
user are issued by finding to which category he
belongs in order to apply the aggregate usage
preferences of previous users in that category. Even
though several categorization techniques have been
proposed, we shall concentrate on Data Clustering
to illustrate DF in our Web service recommender
system. However our approach is independent of the
specific categorization technique used in a DF-based
recommender system.
In DF, clustering is used to create the user
categories mentioned above by considering the set of
all previous users. The objects are users, and each
dimension of the space represents one of their
relevant demographic characteristics. For a given
cluster C, its density represents the number of users
in it and its radius is a measure of how
demographically dissimilar they are. Then, the
historical data on usage behavior or preferences of
each user in C is used to associate with the cluster C
an aggregate buying behavior. In its simplest form,
this aggregate consists of the list of items (Web
services in our case) p
1
, p
2
, ..., p
c
that were
used/purchased or for which positive feedback was
given by users in C. When a new user requires a
recommendation, the recommender system
computes the cluster C to which he is closest, and
then produces a recommendation of the
corresponding list of items.
4 THE WSRS SYSTEM
4.1 Architecture
The architecture of WSRS as depicted in Figure 1
consists of the following components:
Web services providers. The WSRS system
connects to these servers, which we call Web service
providers, to find Web services required by users.
Web services administrator. Its role consists of
organizing the information on each Web service and
its related operations, as well as the input parameters
needed for the execution of the operations. This
information is kept in the Web service database.
Web services database. Web services’
information retrieved from Web service providers
and managed by the Web service administrator are
WEBIST 2007 - International Conference on Web Information Systems and Technologies
286

stored in the WSRS’s Web service database. In
addition, the information on Web services includes
their profiles. Therefore, the Web service database
records are almost of the following format: (name,
description, associated operations, corresponding
execution paths, popularity, average rating).
Figure 1: Architecture of the WSRS system.
User profile database. This database stores all
users’ profiles information like demographic,
preferences, usage behavior, fidelity, etc.
Recommender. The recommender is the key-
component of the WSRS system. It uses the user
profile database and the Web service database to
recommend the user suitable Web services,
according to his demographic data and previous use
of the system, the similarity of his ratings with other
users’ as well as the similarity between the Web
services he used in the past and available ones.
User. The user is any person looking for Web
services. He initially submits her username so her
profile gets loaded from the user profile database.
4.2 Recommendation Process
The recommendation process involves three
components that are described as follows.
The collaborative recommendation (CF-Rec)
of Web services takes the rating matrix, R (this is the
contents of Table 1 without the fidelity column), and
a target user (to whom the recommendations are
made), U, as inputs. It also uses Equation (1) to
compute the similarity between users and Equation
(2) to predict ratings for the current user.
Recommendations are generated as follow.
CF-Rec(Matrix R, User U){
For each user A from the user profile database{
If A has rated Web services in common with U{
Use Eq. 1 to compute similarity between U and A}
For each Web service j that the user has not yet rated{
Use Eq. 2 to compute the predicted rating, P
U,j
, of
U on j}}
Return Web services j with the highest predictions P
U,j
.}
The content-based recommendation (CN-Rec)
of Web services uses the similarity computation
between Web services. The similarity computation is
performed for the operations that compose Web
services. More precisely, two operations are similar
if their names, inputs, outputs and the keywords that
describe them are respectively similar. This requires
a lexical analysis over all the parameters of a given
operation. Each operation parameter is associated
with a similarity index, which is a constant value
indicating the importance of the parameter in the
similarity computation. Let q
name
, q
in
, q
out
, and q
desc
be respectively the similarity indexes for the name,
inputs, outputs, and keywords for a given operation.
It is supposed that q
name
+ q
in
+ q
out
+ q
desc
= 1. The
similarity computation, Compare, between two
strings uses similar_text() function (Oliver, 1993)
and implemented in PHP: Compare(String S
1
, String
S
2
){return similar_text (S
1
, S
2
);}
The general algorithm to compute the similarity
between Web services follows:
Sim (Operation O
1
, Operation O
2
) {
name
Å Compare (O
1
.name, O
2
.name);
in
Å Compare (O
1
.in, O
2
.in);
out
Å Compare (O
1
.out, O
2
.out);
desc
Å Compare (O
1
.desc, O
2
.desc);
Return q
name
·name + q
in
·in + q
out
·out + q
desc
·desc;}
Now, let c be the current operation, v
j
the rating
prediction for operation j, and v
c
the rating for the
current operation. The algorithm for content-based is
as follows:
CN-Rec (Operation c){
For j in the set of operations{
v
j
Å v
c
·Sim(c,j);}
H
Å Operations with highest rating prediction v
j
Return Web services with associated operations
in H}
The demographic recommendation. Using the
user’s demographic profile, π, WSRS computes the
distance between the user’s demographic profile and
the centroid of each cluster, to find the nearest
cluster, C
j
, for a certain j ∈ {1, …, k}, where k is the
number of clusters in the demographic clustering
table. We use Equation 3 to compute the average
fidelity,
)( jf
w
, of a Web service, w, that users in
C
j
have chosen in the past:
Web service
database
User profile
database
Web service
administrator
Recommender
Web service
recommendations
Use
r
Implicit and
explicit
feedback
Web service
p
roviders
Filtering
techniques
WSRS: A WEB SERVICE RECOMMENDER SYSTEM
287

∑
∑
∈
∈
⋅
=
Uu
wu
Cu
wuwu
w
n
fn
jf
j
,
,,
)( (3)
where
wu
f
,
is the fidelity of user u for the Web
service w, and n
u,w
is the number of times user u
used the Web service w. Starting with users in
cluster C
j
, the WSRS system uses Table 1 to select
the fidelity values of these users, and then computes
the average fidelity values (Equation 3).
Table 1: The clustering table: each cluster C
j
is associated
with Web services that present highest average fidelity
values with users in C
j.
Finally, Web services with highest average
fidelity values are selected to be part of the
clustering table (Table 1), updated in the same
manner, as long as Web services are used.
The demographic recommendation (DF-Rec) of
Web services takes Table 1 as input to recommend
Web services that users in C
j
have chosen and used
in the past and that have the highest average fidelity
values. The DF-Rec process follows:
DF-Rec (Clustering table K, U’s demographic profile π){
For each cluster C
j
from K{
d
j
Å distance between π and C
j
}
C
Å cluster corresponding to minimum distance d
j
Return Web services j with the highest usage
percentage that users in C have used in the past}
4.3 Implementation
Currently, we only implemented the CF-Rec and
CN-Rec algorithms. However, an example to
illustrate the execution of these algorithms is
hereafter provided.
Let us illustrate the WSRS system, using the
following scenario: Suppose that John, a first year
student in the department of History is interested in
the conversion between Roman symbols and Integer
representations. For example, converting Integer
“1200” into Roman symbols results in “MCC”. The
WSRS system contains IRomanservice Web service,
which consists of IntToRoman and RomanToInt
operations. However, John does not know that
information. Nonetheless, he decides using WSRS.
John needs to log in before using the system and
receiving recommendations. This allows WSRS to
gather his demographic data, track his activities
(Web services he accesses, how many time, the URL
links he follows, etc.), and update his fidelity.
After login in, John is automatically presented
with a list of Web services recommended by the DF-
Rec part of WSRS. This list is likely to include
IRomanservice, since John is an historian and most
of historian may have used the WSRS system
before. Moreover, John can also input keywords,
such as “convert integer”, to describe his need, and
be presented another list of Web services to use. Let
us suppose that at least one of these two lists include
IRomanservice, John thus uses this Web service
now. More precisely, he clicks on IRomanservice
link and gets access to IntToRoman and RomanToInt
Operations. He then opts, for instance, to use the
former operation. If John tries the operation
IntToRoman with “1200” as the input parameter, he
gets the corresponding roman value, “MCC”.
Meanwhile, the WSRS system update John usage
behavior of IRomanservice by incrementing the
number of times he used this Web service by “1”,
and by re-computing his usage percentage (fidelity)
versus all the other Web services he used in the past
within WSRS (if any). Moreover, John is invited to
provide a rating for the operation InToRoman, and
equivalently for the Web service IRomanservice.
Following the rating step, the CF-Rec and CN-
Rec recommendation processes now take place.
According to John’s previous use of the system (he
may have used and evaluated other Web services in
the past) and his actual feedback (the rating provided
to IntToRoman operation), he receives Collaborative
and Content-based Web service recommendations.
Collaborative recommendations result from the CF-
Rec process, while Content-based recommendations
come from the CN-Rec process. Plus, John is
provided with URL links to other Web services
related to the IntToRoman operation, which may
help him find more information about Web services
of his interest. John then may decide to use other
Web services from the recommendations, or to
follow other URL links.
5 RELATED WORK
Zhang et al. (2002) mentioned several problems
related to search mechanisms in the WSDL/UDDI
systems; the most important one is the lack of
accuracy of the search results, which, for its part, is
particularly due to the lack of Web services
categorization. Therefore, they introduced a new
Clusters Associated Web services
C
1
ConvertRate, ValidateEmail, getQuote
C
2
ValidateEmail, getQuote
… …
C
k
IRomanservice
WEBIST 2007 - International Conference on Web Information Systems and Technologies
288

platform, AUSE (Advanced UDDI Search Engine).
AUSE uses BE4WS (Business to Explore for Web
Services) to facilitate cross research and USML
(UDDI Search Markup Language) to support more
complex requests. In WSRS, the demographic
recommendation process is based on the
categorization of users and/or Web services. We are
still conducting the implementation of this process,
but we hope that to reach a better accuracy than
AUSE/USML.
Limthanmaphon and Zhang (2003) used Case-
Based Reasoning to search for Web services. Wang
et al. (2003) proposed the “query by example”
process. In this case, partial description of the Web
service is provided as an input to the system, which
extracts keywords to compare with textual
information of other Web services. The system
returns Web services having similarity values higher
than a certain threshold. The resulting set of Web
services is then refined with the structure-matching
techniques on WSDL documents.
Contrary to the approaches presented above, the
WSRS system takes into account the user’s profile
to tailor Web service recommendations accordingly.
WSRS provides user with Web services that better
satisfy their needs and requirement because the
recommendation process is based on the implicit and
explicit feedback gathered from the user during his
activities on the WSRS system.
6 CONCLUSION
In this paper, we introduced WSRS, which uses
collaborative, content-based, and demographic
filtering techniques to provide users with
recommended Web services. In a more general
context, the WSRS system can be integrated in any
UDDI to extend its registry with additional
functionalities. This integration is left for future
work. Since creating profiles allow the WSRS
system to track people and get access to which Web
services they are interested in, there is a real need to
introduce privacy-preserving mechanism in the Web
service recommendation process. Moreover, it could
happen that malicious users decide to cheat the
WSRS system with false ratings, with many motives
behind this kind of behavior, such as fun and profit
(Lam, 2004). This practice brings out a dangerous
aspect for the WSRS system (affecting its reputation
for instance). Therefore, we also continue
investigating ways to adequately address this issue.
REFERENCES
Breese J., Heckerman D., and Kadie C. Empirical analysis
of predictive algorithms for collaborative filtering.
Proc. of 14th Conf. on Uncertainty in Artificial
Intelligence (UAI-98). Pages 43–52, Madison, WI,
1998.
Burke, R. Hybrid recommender systems: Survey and
experiments. Customer Modeling and Customer-
Adapted Interaction 4(12):331–370, 2002.
Goldberg D., Nichols D., Oki B.M., and Terry D. Using
collaborative filtering to weave an information
tapestry. Communications of the ACM 35(12):61–70,
1992.
Lam S.K., and Riedl J. Shilling recommender systems for
fun and profit. Proc. of the 13
th
Intern. Conf. on World
Wide Web (WWW ’04). Pages 393–402, New York,
NY, 2004.
Limthanmaphon B., and Zhang Y. Web service
composition with case-based reasoning. Proc. of the
4th Australasian Database Conf. on Database
Technologies. Pages 201–208, Adelaide, Australia,
2003.
Mobasher B. and Anand S.S., Eds. Proc. of the 3rd Intern.
Workshop on Intelligent Techniques for Web
Personalization (ITWP 2005), 19th Intern. Joint Conf.
on Artificial Intelligence (IJCAI 2005). Edinburgh,
Scotland, 2005.
Oliver J. Decision Graphs – An extension of Decision
Trees. Department of Computer Science, Monash
University, Technical Report No. 92/173. Clayton,
Victoria, 1993.
Pazzani M. A framework for collaborative, content-based
and demographic filtering. Artificial Intelligence
Review, 13(5-6):393–408, 1999.
Resnick P., Iacovou N., Sushak M., Bergstrom P., and
Riedl J. Grouplens: An open architecture for
collaborative filtering of netnews. Proc. of the 1994
Computer Supported Collaborative Work Conf. Pages
175–186, North Carolina, 1994.
Schafer J.B., Konstan J.A., and Riedl J. Recommender
systems in e-commerce. EC’99: Proc. of 1st ACM
Conf. on Electronic Commerce. Pages 158–166,
Denver, CO, 1999.
Tsalgatidou A., and Pilioura T. An overview of standards
and related technology in Web Services. Distributed
and Parallel Databases, 12(3): 135–162, 2002.
Turban E., King D., Viehland D., and Lee J. Electronic
Commerce: A Managerial Perspective. Prentice Hall,
2006.
Wang Y., and Stroulia E. Semantic Structure Matching for
Assessing Web-Service Similarity. Proc. of the 1rst
Intern Conference on Service Oriented Computing.
Trento, Italy, 2003.
Zhang L., Li H., Chang H., and Chao T. XML-based
advanced UDDI search mechanism for B2B
integration. Proc. of the 4th IEEE Int’l Workshop on
Advanced Issues of E-Commerce and Web-Based
Information Systems (WECWIS 2002). Pages 1530–
1534, Newport Beach, CA, 2002.
WSRS: A WEB SERVICE RECOMMENDER SYSTEM
289