
by lifting the data to additional dimensions, but its
computation complexity is high.
Facing plenty of services, it is necessary to recommend
component services to a new composite service according to
similar composite services. The similar composite services
may have similar aims, approaches or functions, even similar
QoS as they have similar component services. This feature
rapidly makes excellent component services popular, once
they are selected in the beginning by several composite
services. The composite services and component services
form a two-dimension matrix, and the element of the matrix
indicates one component service’s utility to a composite
service. Because we do not know how well some component
services are composed to a composite service, the matrix is
sparse and has missing data. The methods of matrix factoring
have proven to be effective for recommendation of user
preference in the contest of Netflix. SVD (Singular Value
Decomposition), one of these methods, can derive the latent
semantic information from partial information to predict the
missing information.
To solve these problems, we proposed a method, called
CRE (Clustering and Recommendation for OWL-S Web
services in Evolution). In section 2, we reviewed the
previous work in service clustering and recommendation. In
section 3, we cluster Web services according to their
similarities, and the outputs of the clusters construct two
subspaces (component service-topic subspace and composite
service-topic subspace). In section 4, a model called service
space was constructed for service recommendation, and a
service recommendation method, based on matrix
decomposition, was proposed. In section 5, experiments
about clustering and matrix decomposition are conducted.
The contributions of this paper are:
First, we combine TF/IDF and ontology to compute more
precisely the similarity of Web services. Based on the
similarity we use Agglomerative clustering method to
classify the services.
Second, considering four factors, we build a service
space model, by which we propose a service
recommendation method.
II. R
ELATED WORK
With the recent success of recommendation systems in e-
commerce, like Amazon, and online content distribution, like
Netflix, analogous recommendation techniques have been
proposed to recommend services [3, 4]. As the key
technology of recommendation systems, Collaborative
Filtering (CF) [3] considers user disparity and uses
experience of one group of users to predict experience of
another similar group of users, who actually have not
invoked the services yet. Current CF techniques generally
can be divided into two classes: neighborhood-based and
model-based. The former needs to identify users who share
similar QoS experience with new users. The number of
services invoked by each user is limited, thus identifying the
number of users who have invoked the same services is
usually difficult. Consequently, the QoS data sparsity issue
hinders the neighborhood-based approaches. Model-based
approaches build a global model according to the observed
QoS data, which can be used to predict QoS values.
Nevertheless, model-based approaches need to build models
in advance, and some models have difficulty explaining
results of predictions.
As the number of Web services grows fast, it is more
difficult to discover services that satisfy both functional and
non-functional users’ requirements. Existing techniques that
can be applied to service discovery fall into two broad
categories: functionality-based and QoS-based.
Functionality-based service discovery is to discover services
that satisfy user-required functionality. Information retrieval
techniques and semantic technologies have been used to
improve the accuracy of service discovery. QoS-based
approaches are used to differentiate services based on their
QoS performance [5]. [6-8] propose user-based CF
algorithms, and they predict QoS by assuming that similar or
trustable users tend to receive similar QoS from similar
services. A hybrid algorithm was developed, which enhances
the user-based approach by integrating item-based CF to
obtain better QoS prediction accuracy [7]. Thereafter, many
similar algorithms have been developed. They use additional
information, such as users’ locations [9], invocation
frequencies of services, and users’ query histories to improve
the QoS recommendation. Both user and item based
approaches adopt the neighborhood centric strategy in CF,
which searches the local neighborhood to find similar users
or for recommendation.
Recently, a model based CF algorithm was proposed and
obtained higher prediction accuracy [10]. First, it uses the
user-based approach to find top-k similar users. Second,
matrix factorization is employed to construct a global model
according to the user neighborhood information. Services
prediction can be modeled as a general matrix completion
problem, which has applications in many science and
engineering domains. It presumes that the matrix has a low-
rank or approximately low-rank structure. However, rank
minimization is NP-hard because the rank function is non-
convex and discontinuous [11, 12]. It has been demonstrated
that the nuclear norm provides the tightest convex relaxation
of the rank minimization problem [13]. Many optimization
algorithms have been developed recently, which apply
nuclear norm as a convex surrogate to effectively tackle the
matrix completion problem.
Zhang [14] etc. proposed an approach which combines
social network and collaborative filtering techniques in a
unified framework to predict the missing QoS values of
manufacturing services for an active service user. Their
method alleviates the data sparsity and the cold start
problems that hinder the traditional collaborative filtering
techniques. Sun [15] etc. presented a new similarity measure
for web service similarity computation and proposed a novel
collaborative filtering approach, called normal recovery
collaborative filtering, for personalized web service
recommendation. They conducted large-scale real-world
experiments, and the results showed their approach achieves
better accuracy than other competing approaches. Cao [16]
etc. designed a cube model to describe the relationship
665665665665