A Proposal to Maintain the Semantic Balance in Cluster-based Data
Integration Systems
Edemberg Rocha Silva
1
, Bernadette Farias Lóscio
2
and Ana Carolina Salgado
2
1
Federal Institute of Education, Science and Tecnhology of Paraíba, João Pessoa, Brazil
2
Federal University of Pernambuco, Recife, Brazil
Keywords: Semantic Balance, Semantic Clusters, Dynamic Data Integration Systems, Schema Evolution, Clustering
Measure.
Abstract: With the large volume of data sources on the Web, we need a system that integrates them, so that the user
can query them transparently. For efficiency in queries, integration systems can group these sources in
clusters according to the semantic similarity of their schemas. However, the sources have autonomy to
evolve their schema, and to join or to leave the integration system at any time. This autonomy may cause a
problem which we define as semantic unbalance of clusters. The semantic unbalance can compromise the
formation of clusters and hence the efficiency of the submitted queries. In this paper, we propose a solution
to the semantic balance of clusters in dynamic data integration systems based on self-organization. We also
introduce a measure to evaluate how much the clusters are semantically unbalanced.
1 INTRODUCTION
The increasing number of distributed, autonomous
and heterogeneous data sources, (ie: XML
documents, relational database and HTML pages
among others), has motivated the need for data
integration systems, allowing users to query those
sources transparently (Roth and Skritek., 2013; Pires
et al., 2012). In this sense, dynamic data integration
systems, as Peer Data Management Systems
(PDMS) (Halevy et al., 2006), Grid (Zamboulis et
al., 2010) or systems based on a pay-as-you-go
strategy (Halevy et al., 2008), have been used to
improve the data sharing of data sources distributed
on the Web or on the cloud (Wall and Angryk,
2011). Some of these systems have a dynamic
behaviour, i.e., their data sources have autonomy to
join/leave the system and to change their own
schemas. By convention, we will call these dynamic
data integration systems as data integration systems
and their data sources as peers.
Aiming to reduce the search space for queries,
their response time and to diminish the message
traffic on the network, some data integration systems
organize their networks in clusters (peer grouping)
(Raftopoulou and Petrakis, 2008; Montanelli et al.,
2011; Kantere et al., 2008). According to
Raftopoulou and Petrakis (2008), the query
processing in a data integration environment can be
improved if peers are grouped. In order to group
peers, semantic similarity measures (Montanelli et
al., 2011; Pires et al., 2012) between peer schemas
may be used. Once these peers are grouped, a query
can be sent to the cluster that may offer the best
answer. Data integration systems like ESTEEM
(Montanelli et al., 2011), SPEED (Pires et al., 2012)
and OntoZilla (Joung and Chuang, 2009) create their
clusters according to the semantic similarity of their
schemas. In these systems, peers' schemas are
represented by a conceptual ontology, called peer
ontology. In addition, some systems have a
conceptual ontology that represents the cluster
schema, which is called cluster ontology.
Some of the clusters-based systems have two
levels of connections between clusters: the intra-
cluster and the inter-cluster level (Ayyasamy and
Sivanandam, 2010). Each of these levels has an
overlay network whose connections are established
through the semantic similarity of the respective data
source schemas. The intra-cluster level comprises
the connections between peers that belong to the
same cluster. In this level, a connection between
peers is established only if there is a minimum
semantic similarity between the peer ontology and
the cluster ontology. On the other hand, the
connection at the inter-cluster level comprises the
90
Rocha Silva E., Farias Lóscio B. and Carolina Salgado A..
A Proposal to Maintain the Semantic Balance in Cluster-based Data Integration Systems.
DOI: 10.5220/0004897000900098
In Proceedings of the 16th International Conference on Enterprise Information Systems (ICEIS-2014), pages 90-98
ISBN: 978-989-758-027-7
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
connections between different clusters. In this case,
a minimum semantic similarity between clusters’
ontology is required. Then, clusters that are
connected to other clusters are called semantic
neighbors.
The dynamic behaviour of peers in this type of
data integration systems may cause evolutions of
clusters’ ontologies. These evolutions can result in a
situation in which a cluster has peers and/or
neighbors with low semantic similarity. When this
situation occurs, we say that the cluster is
semantically unbalanced. This unbalance can occur
both at the intra and at the inter-cluster level.
A semantic unbalance may cause an undesirable
behaviour during query processing on cluster-based
data integration systems. At the intra-cluster level,
with the existence of peers with low semantic
similarity with the cluster, these peers will probably
not contribute to respond expressively the queries
that arrive to the cluster. These peers could
contribute more to other clusters with which they
have more semantic similarity.
As the inter-cluster connections are established
by semantic similarity between clusters, when a
query is submitted it is routed through inter-cluster
connections. When semantically unbalanced clusters
exist at the inter-cluster level, a query can be routed
to clusters which do not offer meaningful results for
this query. Furthermore, this situation will contribute
to the degradation of query response time.
We propose, in this paper, a solution to the
semantic unbalance problem, which is based on a
self-organization approach, i.e., without human
intervention. This paper examines the types of
dynamic behaviours (peer joining/leaving or peer
schema evolution) that may cause the semantic
unbalance of clusters, and proposes a solution that
automatically produces semantically balanced
clusters. To evaluate the network organization
before and after actions of semantic balance, we use
the measure called clustering efficiency
(Raftopoulou and Petrakis, 2008). Furthermore, we
introduce a measure that quantifies the level of
semantic balance of all the network of peers, called
semantic balance coefficient.
The remainder of this paper is organized as
follows. Section 2 discusses the environment that is
used to illustrate the semantic unbalance problem.
The schema evolution in data integration systems
will be discussed in Section 3. Section 4 discusses
how this evolution may cause semantic unbalanced
clusters. Section 5, describes our solution based on
self-organization (Conforti et al., 2004; Pires et al.,
2012) (without human intervention) for the semantic
unbalance problem. Section 6 presents measures that
evaluate the network organization. Our experiments
and results are described in Section 7. Related works
and concluding remarks are in Sections 8 and 9,
respectively.
2 THE SPEED SYSTEM
In this section, we present a semantic cluster-based
dynamic data integration system called SPEED
(Semantic PEEr Data Management System) (Pires et
al., 2012). In SPEED, the peers represent their
schemas through ontologies. In addition, each
cluster has its own cluster ontology which describes
the schemas of their peers. The cluster is managed
by a peer called super-peer. The super-peer is a peer
belonging to the cluster, which has more processing
capacity and good network connectivity. It is
responsible for managing query processing and data
integration.
The cluster ontology is stored in the super-peer.
In SPEED, the semantic similarity between peers is
computed using a tool called SemMatcher (Pires et
al., 2012). This tool performs a semantic-based
ontology matching process that considers, besides
the traditional terminological and structural
matching techniques, a semantic-based one. The
matching process produces a set of semantic
correspondences (alignments) and a Global
Similarity Measure (GSM) between the two matched
ontologies. The GSM is calculated considering the
semantic similarity measure of the generated
correspondences.
For a peer to join a cluster, the GSM between the
peer ontology and the cluster ontology should be
greater than or equal to the threshold of minimum
semantic similarity of the cluster, called cluster
threshold. However, if this similarity is a value
below the cluster threshold, the peer will find
another semantic similar cluster or form a new one.
This new cluster will be a neighbor of another
cluster if the semantic similarity between their
cluster ontologies is greater than or equal to the
minimum neighborhood threshold, called neighbor
threshold.
When a peer joins a cluster, a merge between its
peer ontology and the cluster ontology is performed
to generate a new cluster ontology. The SPEED uses
the OntMerger tool (Pires et al., 2012) to perform
the merge between ontologies. The OntMerger tool
takes as arguments two ontologies (i.e, the cluster
ontology and the peer ontology) and the set of
correspondences between them (generated by
AProposaltoMaintaintheSemanticBalanceinCluster-basedDataIntegrationSystems
91
SemMatcher). As a result, the tool produces a new
version of the cluster ontology containing the
elements of both input ontologies as well as the
semantic correspondences between the new cluster
ontology and the peer ontology.
3 SCHEMA EVOLUTION IN
DYNAMIC DATA
INTEGRATION SYSTEMS
As stated by (Curino et al., 2013), the schema
evolution is a change likely to happen over the life
cycle of the systems, but it needs to be considered so
as not to affect the queries. Actions should be taken
so that the queries are not harmed, by having empty
results or different results from what was expected
(Genevès et al., 2011).
Once the data integration systems are composed
by data sources, the schema evolution of these
sources may compromise some of the functionalities
of the integration system (such as queries,
composition of semantic clusters, among others).
When there is a schema that represents a cluster
(cluster ontology), based on the peers’ schemas, the
concern is even worst because the cluster schema
should describe the real schemas of the peers within
the cluster. Furthermore, all intra and inter-cluster
connections are established based on clusters’
schemas. The cluster ontology can evolve in three
situations, as described below:
i) when a peer joins a cluster: in this case, a
merge between the current cluster ontology and the
peer ontology is performed, evolving the cluster
ontology.
ii) when a peer leaves a cluster: a peer may leave
the cluster for some reasons: problems in the
physical network, the peer wants to leave the
integration system or it no longer has a minimum
semantic similarity with the cluster. When a peer
leaves a cluster, the schema elements that belong
exclusively to this peer should be removed from the
cluster ontology causing the evolution of the cluster
schema.
iii) when a peer schema evolves (evolution of the
peer ontology): the evolutions are related to changes
in the logical definition of the peer data source (add,
delete, reset, or split columns, change the cardinality
of the relationship from 1: N to M: N, among others)
(Sockut and Iyer, 2011). These evolutions should be
also reflected in the cluster ontology.
In all described situations, the cluster schema
evolution may cause problems in the peer clustering,
which will be described in next section.
4 SEMANTIC UNBALANCE
The evolution that occurs in the clusters’ ontologies
can let the clusters on a state in which some intra-
cluster and inter-cluster semantic connections
become below the established thresholds. When this
occurs, we say that there is a semantic unbalance at
the intra-cluster and/or at the inter-cluster levels. We
present the definitions related to the unbalance
problem as follows.
4.1 Intra-cluster Semantic Unbalance
Let C be a cluster and Sp = {p
1
, p
2
, p
3
, p
4
,..., p
n
} the
set of peers belonging to C. C is semantically
unbalanced at the intra-cluster level if there is, at
least, one peer p
i
Sp; in such a way that the GSM
between O(p
i
) and O(C) is below the cluster
threshold, denoted by
. Thus, we can define the
intra-cluster semantic unbalance in the following
way:
p
i
Sp, p
i
| GSM(O(p
i
), O(C)) <
, C is
semantically unbalanced at the intra-cluster level.
There are two actions that can cause an intra-
cluster semantic unbalance: either by a peer joining
the cluster or by the evolution of a peer ontology.
Although a peer leaving the cluster provokes an
evolution of the cluster ontology, it does not result in
an intra-cluster unbalance. When a peer leaves the
cluster the exclusive elements of its peer ontology
will be removed from cluster ontology. Therefore,
the semantic similarity between others peers’
ontologies and the cluster ontology can remain the
same or it can be increased.
We will present an example in order to illustrate
how the evolution of the cluster ontology may cause
a semantic unbalance. Consider that the system
wants to integrate five peers p
1
, p
2
, p
3
, p
4
and p
5
. The
peers’ schemas are represented by ontologies and are
part of the same domain (for example, Education
domain). Suppose that these peers will form a single
cluster C, i.e., each peer ontology has a semantic
similarity with C cluster ontology greater than or
equal to the cluster threshold (
). We assume that
value of
is equal to 0.7. The C ontology is denoted
by O(C) and the p
i
ontology by O(p
i
). The p
1
will be
the first peer to enter C and O(C) will be equal to
O(p
1
). In the sequence, the others peers will join C
as summarized in Table 1. The table columns mean:
Peer: the peer which is joining the cluster C.
ICEIS2014-16thInternationalConferenceonEnterpriseInformationSystems
92
Schema Evolution: the way as the C ontology
evolved. The OntMerger tool was used to
performer the merge between cluster ontology
and peer ontology.
GSM: the GSM between O(C) and O(
p
i
), when p
i
is joining the cluster. The GSM values were
computed by the SemMatcher tool.
When p2 joins C, the merge between O(C) and
O(p
2
) is performed, evolving the O(C). After, the
similarity measures are recomputed. We can observe
the GSM between O(C) and O(p
1
) decreased from 1
to 0.84, after p
2
joins the cluster. However, all the
GSM remain above
. Similar to p
2
join, when p
3
, p
4
and p
5
join the cluster C, the O(C) also evolves.
According to Table 1, after p
5
has joined the cluster,
Table 1: Example of semantic unbalance due evolution of
cluster ontology.
Peer Schema Evolution GSM
p
1
O(C)=O(p
1
) p
1
= 1.0
p
2
merge(O(p
2
),O(C))
p
1
=0.84
p
2
=0.92
p
3
merge(O(p
3
),O(C))
p
1
=0.75
p
2
=0.85
p
3
=0.85
p
4
merge(O(p
4
),O(C))
p
1
=0.72
p
2
=0.81
p
3
=0.81
p
4
=0.79
p
5
merge(O(p
5
),O(C))
p
1
=0.60
p
2
=0.71
p
3
=0.70
p
4
=0.69
p
5
=0.85
there were two intra-cluster semantic unbalances
within C. We can observe that GSM(O(p
1
),O(C))
and GSM(O(p
4
),O(C)) decreased from 0.72 and 0.81
to 0.60 and 0.69, respectively (below
). These two
peers are provoking the semantic unbalance of C.
4.2 Inter-cluster Semantic Unbalance
Let C be a cluster and Nc = {v
1
, v
2
, v
3
, v
4
,..., v
k
} the
set of neighbor clusters of C. C is semantically
unbalanced at the inter-cluster level if there is at
least a neighbor cluster v
j
Nc, such a way that the
GSM between O(v
j
) and O(C) is below the neighbor
threshold, denoted by
. In this way, we can define
the inter-cluster semantic unbalance in the following
way:
v
i
Nc, v
i
| GSM(O(v
i
), O(C)) <
, C is
semantically unbalanced at the inter-cluster level.
The evolutions occurred in the cluster ontologies
can cause a semantic unbalance at the inter-cluster
level. As the inter-cluster level is established
according to the GSM computed between cluster
ontologies, if at least one of these ontologies
changes, the semantic similarity between them can
decrease for a value below
.
To understand how changes in the cluster
schema may cause inter-cluster semantic unbalance,
let’s consider C formation according to the Table 1.
After p
5
, the peer p
6
joined the system and the GSM
between O(C) and O(p
6
) is equal to 0.4. This value
is below to
, thus p
6
will form its own cluster (C’).
Considering
equals 0.4, C and C’ will become
semantic neighbors. Suppose that in other time, the
peer p
7
joined the system and the values of
GSM(O(C),O(p
7
)) and GSM(O(C’),O(p
7
)) are 0.53
and 0.83, respectively. Thus, p
7
will join C’,
evolving O(C’). After p
7
joined, the value of
GSM(O(C),O(C’)) decrease from 0.4 to 0.37, value
above the
. Therefore, the join of p
7
caused an
inter-cluster semantic unbalance.
5 ACTIONS TO SEMANTIC
BALANCE
In this paper, we propose a self-organization based
solution for the previously described semantic
unbalance problem. In our proposal, each cluster is
able to self-organize in order to keep the semantic
balance at the intra-cluster level as well as at the
inter-cluster level, with no need of human
intervention.
In a general way, the solution for the semantic
unbalance problem consists of periodically checking
if there is a semantic unbalance both at the inter-
cluster and at the intra-cluster level. This can be
done recalculating the semantic similarity measures
between the cluster ontology and the peers’
ontologies, as well as between clusters’ ontologies.
The cluster could check if there is a semantic
unbalance when the actions that can provoke the
evolution of cluster ontology are detected. When this
occurs, all the GSM must be recomputed and the
cluster checks whether they are below the
thresholds. This action may cause an overload in the
cluster due to the dynamic nature of this type of data
integration system. To avoid this overload, we will
perform, periodically, the detection of the semantic
unbalance situation. This period can be set in the
system. Figure 1 illustrates the semantic balance
algorithm for the intra-cluster and inter-cluster
levels.
AProposaltoMaintaintheSemanticBalanceinCluster-basedDataIntegrationSystems
93
Let C
i
be a cluster and Sp = {p
1
, p
2
, p
3
, p
4
,..., p
n
} the
set of peers belonging to C
i
(lines 9 and 10). C
i
has
the super-peer s
i
(line 11) and the set of semantic
neighbors S
V
= {v
1
,..., v
m
} (line 12). Initially, the
semantic balance solution tries to repair the
problems at the intra-cluster level, given that the
evolution of the cluster ontology, due to semantic
balancing actions in this level, can cause a semantic
unbalance at the inter-cluster level. However, the
balance actions at the inter-cluster level do not
impact the intra-cluster level because the cluster
ontology remains unchanged with the input or output
of new neighbors.
At the intra-cluster level, each unbalanced peer p
i
will be inserted into a list L
P
in ascendant order,
according to the GSM between O(p
i
) and O(C
i
)
(lines 15 to 20). If the super-peer is an unbalanced
peer (line 22), a new super-peer among the other
peers that are not unbalanced (lines 24 and 25). This
new super-peer will store O(C
i
) (line 26). While L
p
is not empty (line 30), for each peer p
k
removed
from L
P
(lines 32 and 33), s
i
sends a message, to
each neighbor v
j
, to search for a new cluster to
receive p
k
(lines 36 and 37). The GSM of the
remaining peers of L
P
are recalculated (line 34)
because with the output of p
i
, it is likely that some
peers in this list have become balanced, as discussed
in Section 4.1. The message contains the O(p
i
). The
search will end when a particular TTL (Time-to -
Live), is achieved (line 39). Only the clusters that
have greater semantic similarity and capacity to
absorb the load generated by p
k
will send back to s
i
a
message of interest by the peer (lines 41 and 42).
Each interested cluster is added into a list S
C
(line
43).
1.
Algorithm Semantic Balance()
2. {
3.
,
:Threshold;
4. v,C
i
: Cluster;
5. S
C
, S
V
: Set of Clusters;
6. p
i
,p
k
,s
i
: Peer;
7. S
P
,L
P
: Set of Peers;
8.
9. C
i
← retrieve current cluster;
10 S
P
← retrieve the peers of C
i
;
11 s
i
←retrieve current super-
peer;
12 S
V
← retrieve neighbors of C
i
;
13
14
L
P
← ;
15 FOR EACH p
i
IN S
P
DO
16
IF (GSM(O(p
i
),O(C
i
))<
) THEN
17 {
18 add p
i
into L
P
in ascendant
1
9 order by GSM(O(p
i
),O(C
i
));
2
0 }
2
1
2
2
IF s
i
L
P
THEN
2
3 {
2
4 s
i
← select new super-peer
2
5 from {S
P
– L
P
};
2
6 store O(C
i
) into s
i
;
2
7 }
2
8
2
9
S
C
← ;
3
0
WHILE (L
P
≠ ) DO
3
1 {
3
2 p
k
← remove first element
3
3 from L
P
;
3
4
IF (GSM(O(p
k
), O(C
i
))<
)
THEN
3
5 {
3
6 FOR EACH v
j
IN S
V
DO
3
7 send O(p
k
)to v
j
;
3
8 }
3
9 WHILE (TTL ≠ 0)
4
0 {
4
1 C
Z
← receive interested
4
2 cluster;
4
3 add C
Z
into S
C
;
4
4 }
4
5
IF (S
C
≠ ) THEN
4
6 {
4
7 C
Z
← select cluster from
4
8 S
C
with greater
4
9 GSM(O(p
k
), O(C
i
);
5
0 unmerge(O(p
k
),O(C
i
));
5
1 disconnect p
k
from C
i
;
5
2 connect p
k
to C
Z
;
5
3 }
5
4 ELSE
5
5 create new cluster with
p
k
;
5
6 }
5
7
5
8 FOR EACH v
j
IN S
V
DO
5
9 {
6
0
IF (GSM(O(v
j
),O(C
i
))<
) THEN
6
1 disconnect v
j
from C
i
;
6
2 }
6
3}
Figure 1.
After receiving all messages of interest, s
i
will
choose from S
C
, the cluster C
z
which has the highest
semantic similarity to receive p
k
(lines 47 to 49). s
i
will run the “unmerge” between the O(C
i
) and O(p
k
)
(line 50) and disconnect p
k
from itself (line 51). The
“unmerge” is inverse to the merge operation and it
ICEIS2014-16thInternationalConferenceonEnterpriseInformationSystems
94
removes the schema elements that exclusively
belong to O(p
k
) from O(C
i
). Afterwards, p
i
will
connect to C
z
(line 52). If no cluster shows interest
by p
i
, this peer will form its own cluster, according
to the process of cluster formation established by the
system (lines 54 and 55).
Once the actions of intra-cluster balance are
finished, it is time to check whether there is an inter-
cluster semantic unbalance and then balance them.
During the inter-cluster balancing, the GSM for each
v
j
in S
V
is recomputed (lines 58 to 60). Each
unbalanced v
j
is disconnected from C
i
(line 61).
However, C
i
periodically sends a message in the
network to search for new semantic neighbors. The
clusters that have minimum semantic similarity with
O(C
i
) will be connected to C
i
.
6 MEASURING CLUSTERING
QUALITY
We introduce the semantic balance coefficient
as
a measure that quantifies the level of semantic
balance of the whole network. To evaluate the
network organization, i.e., the set of clusters that
composed the system, before and after the semantic
balance actions, we present a measure called
clustering efficiency
k
(Raftopoulou and Petrakis,
2008) that quantifies network organization by
exploiting the underlying network structure.
6.1 Clustering Efficiency
To measure the efficiency of the network
organization, we use the clustering efficiency
measure
k
(Raftopoulou and Petrakis 2008), which
is defined based on the set of clustering efficiency
measures for the clusters that compose the network.
Formally, the clustering efficiency k
i
for a cluster C
i
is defined as the number of peers p
j
semantically
similar to C
i
(sim(O(C
i
),O(p
j
))>
) that can be
reached from C
i
within a TTL, divided by the total
number of peers p
k
in the network similar to C
i
. The
dg() function is the distance (measured in number of
hops) of peers in the network.
1
1
:{ ( , ) , ( ( ), ( ) }
:{ ( ( ), ( ))}
N
j
N
k
pj dg Ci pj TTL sim O Ci O pj
ki
pk sim O Ci O pk
(1)
The clustering efficiency for the network as a
whole
k
is defined as the clustering efficiency
average (over all clusters in the network) . We use
the clustering efficiency to evaluate the peer
clustering after and before the semantic balance
actions.
N
i
ki
N
k
1
1
(2)
The clustering efficiency measure gives
information about the underlying overlay network,
and looks at how the network is structured at a larger
scale. Clustering efficiency measure will have values
in the interval [0;1]. According to (Raftopoulou and
Petrakis, 2008), the highest the value of clustering
efficiency is, the better the underlying network
organization is.
6.2 Semantic Balance Coefficient
To evaluate how much the network is semantically
balanced, we introduce the semantic balance
coefficient measure based on recall measure
(Rijsbergen, 1979). We propose this measure
because the clustering efficiency measure doesn’t
consider the intra-cluster and inter-cluster
connections that are semantically unbalanced. For
each cluster we considered the number of intra and
inter-cluster connections that are semantically
balanced over the total number of cluster
connections.
First, we defined the intra-cluster semantic
balance coefficient for a cluster C
i
as the number
of peers p
i
belonging to C
i
and whose
GSM(O(C
i
),O(p
i
)) is greater than or equals
,
divided by the total of intra-cluster connections n in
C
i
, i.e., the number of peers within C
i
:
=
1
1
:{ , ( ( ), ( )) }
n
i
pi pi Ci GSM O Ci O pi
n
(3)
The measure is equal to one if the whole set of
peers p
i
of C
i
is semantic balanced. Conversely, the
measure is equal to zero if all peers of C
i
are
semantic unbalanced.
Similar to the , the inter-cluster semantic
balance coefficient
for a cluster C
i
is defined as the
number of C
i
neighbors (v
i
)whose GSM (O(C
i
),
O(v
i
)) is greater than or equals
, divided by the total
of inter-cluster connections m in C
i
, i.e., the number
of neighbors of C
i
.
1
1
:{ , ( ( ), ( )) }
m
j
vi vi Ci GSM O Ci O vi
m
(4)
The
measure is equal to one if all neighbors of
C
i
are semantically balanced at the inter-cluster
level. Conversely, the
is equal to zero if all
AProposaltoMaintaintheSemanticBalanceinCluster-basedDataIntegrationSystems
95
neighbors of C
i
are semantically unbalanced at the
inter-cluster level. So, the semantic balance
coefficient
i
for a cluster C
i
is the ratio between its
and
coefficients and its number of intra and
inter-cluster connections:
mn
i
(5)
The semantic balance coefficient for all the network
is the ratio between the sum of all
i
and the
number of clusters N in the network.
N
k
i
N
1
1
(6)
The semantic balance coefficient is a measure that
gives information about how the network is
semantically balanced at a large scale. This measure
is used to evaluate our solution for the semantic
unbalance problem, as we discuss in the next
section.
7 EVALUATION
In this section, we present the results of some tests
performed to evaluate our proposal. The tests were
performed using the SPEED prototype. The main
goal of the proposed experiment is initially to create
some clusters and in the next step to evaluate the
network organization when one or more clusters
become unbalanced. To evaluate our proposal, we
consider two different scenarios: in the first one, our
proposal was used to solve the semantic unbalancing
problem and in the second scenario our proposal was
not considered.
Both the prototype and our solution to the
semantic balance were implemented in Java
language. Our experiments were done on a single
CPU (Core 2 Duo E8400, 3Ghz), simulating forty
five (45) peers. Issues about performance were
evaluated and discussed in (Silva et al., 2013).
7.1 Experimental Set-up
Our experiments have been performed in a set of
peer ontologies of the Education domain. Each peer
ontology is represented in OWL (OWL, 2013) and
has about six concepts on average. The cluster
threshold was set to
=0.7 and the neighbor
threshold to
=0.4.
Five different scenarios were generated with
different orders of new peer joins, forming twenty
one clusters on average. In each scenario some peers
were randomly chosen to evolve their schemas. At
each time interval of 200K milliseconds, the clusters
check whether there are semantic unbalances. To
guarantee that all clusters were visited in the
semantic balance solution, we set the TTL to the
number of peers, i.e., equals 45.
7.2 Experimental Validation
Figures 2 and 3 illustrate an overview of the
network, considering the measures
and
k
. The
measures were computed two times for the
situations: (i) without, and (ii) with semantic balance
actions. Therefore, each point on the graphics is the
average value of five scenarios to both situations.
Figure 2 illustrates the variations of clustering
efficiency
k
for each peer join operation, in the two
situations. The graphics pointed out that when the
semantic balance actions are performed,
k
has better
results. In situations (i) and (ii) (Figure 2) the values
of
k
decreased with the join of some peers. These
decrease are justified by the dynamic behavior of
system that allowed new clusters, semantically
balanced, to be formed with the join of new peers in
the system.
Figure 2: Clustering efficiency as a function of peers
joining the network with or without semantic balance
actions.
Figure 3 illustrates how semantic balance
coefficient varies over time, in the two situations. In
situation (i) the coefficient has worst value. After
9x200K milliseconds the coefficient was equal to
0.5, i.e., fifty percent of the connections in the
network were semantically unbalanced. In situation
(i) the values of
increased in the time intervals
2x200K<t<3x200K and 5x200K<t<8x200K. These
increase are justified by the dynamic behavior of the
network that allows new clusters, semantically
balanced, to be formed with new peers joining the
ICEIS2014-16thInternationalConferenceonEnterpriseInformationSystems
96
system. However, the other clusters remained
unbalanced.
Figure 3: Semantic balance coefficient as a function of
time with or without semantic balance actions.
Also in Figure 3, we observe that in situation (ii)
the values of
remain above the values found in (i).
Over time, the semantic balance actions self-
organize the network as close as possible to a
balanced condition, i.e.,
close to value 1. In
situation (ii) (Figure 3) the values of
decreased in
the time intervals 1x200K<t<2x200K,
3x200K<t<5x200K, and 8x200K<t<10x200K. This
situation is also justified by the system dynamic
behaviour that allows new peers to join/leave the
system or peers to evolve their schema, provoking
new semantic unbalances.
The experiments above showed that the use of
our approach for semantic balancing achieves better
values of clustering efficiency and semantic balance
coefficient. Even with the system dynamic
behaviour, our solution presents better results. The
semantic balance actions provoked a network
reorganization, but the clustering efficiency measure
indicated better peer clustering after this actions.
However, the issues related to cost of cluster
reorganization (like complexities in time and
message traffic) will be precisely analyzed in the
future works.
8 RELATED WORK
In this section, we make a brief summary of some
works similar to ours, which intend to maintain the
semantic balance of their clusters.
In (Montanelli et al., 2011), a peer forms a
cluster (founder peer) based on its schema. The
schema that represents the cluster is the one related
to the peer that created it. In this work, only the
intra-cluster level is established. The peer
joining/leaving or the schema evolution of the other
peers do not make the cluster schema to evolve.
Only the schema of the founder peer will evolve and
it represents the cluster. New peers can be found and
added to the cluster by means of probe message that
is submitted to the network. If cluster finds new
peers more similar than the others already belonging
to the cluster, it will replace the less similar peer by
the more similar new one.
Kantere et al. (2008) developed an algorithm for
cluster peers considering a minimum value of the
similarity threshold. There is a schema that
represents the cluster and it is obtained through
merging operations of the schemas of the sources.
The algorithm is able to detect the peers of the
clusters that had its semantic similarity decremented
due to the evolution of the cluster schema. As only
connections at the intra-cluster levels are
established, Kantere et al. (2008) solve the semantic
unbalance only at this level.
In (Raftopoulou and Petrakis, 2008), the cluster
discards the less similar peers by the more similar
ones found in the network. Each cluster starts a
process of reconnection of the more similar peers
every time there is an intra-cluster semantic
unbalance. The reconnection procedure finds the less
similar peers and substitutes them by the more
similar ones according to an established cluster
threshold. There is no inter-cluster level of
connections.
Conforti et al. (2004) describe a set of formation
of clusters following a super-peer topology. Each
super-peer owns a list containing the peers’ schemas
connected to it and it integrates them, i.e., it creates
a super-peer schema (representing the general view
of the cluster schema). By means of a self-
organization process, the peers are grouped in
clusters according to the semantic similarity between
the peers’ schemas. Inter-cluster connections are
established between semantically similar clusters,
considering the super-peers’ schemas. Nevertheless,
there is no procedure that identifies the semantic
unbalance and that fulfils the necessary semantic
balance at the intra-cluster and inter-cluster levels.
Compared to the works discussed above, our
algorithm carries out the semantic balance both at
the intra-cluster and at the inter-cluster levels.
Furthermore, it can be used in a super-peer topology
or not. Any schema evolution that causes the
semantic unbalance is also considered.
AProposaltoMaintaintheSemanticBalanceinCluster-basedDataIntegrationSystems
97
9 CONCLUSIONS AND FUTURE
WORKS
In this paper, we propose a solution to solve the
problem of semantic unbalance of clusters at the
intra and at the inter-cluster levels in dynamic data
integration environments. Our proposal ensures that
the clusters will always have semantically similar
peers according to the established threshold for both
intra-cluster and inter-cluster connections. The
experimental results demonstrated that our solution
is able to detect a semantic unbalance problem and
to perform the corresponding balance actions in
order to keep the semantic balance of the systems
clusters. We are currently improving the solution to
reduce the overload, over the cluster, when verifying
the semantic unbalance.
REFERENCES
Ayyasamy, S., and Sivanandam, S., 2010. A Cluster Based
Replication Architecture for Load Balancing in Peer-
to-Peer Content Distribution. International Journal of
Computer Networks & Communications (IJCNC),
vol.2, pp. 158-172.
Conforti, G., Ghelli, G., Manghi, P., and Sartiani, C.,
2004. A Self-organizing XML P2P Database System.
Proceedings of the 2004 international conference on
Current Trends in Database Technology, pp. 456-465.
Curino, C., Moon, H. J., D., Alin, and Zaniolo, C., 2013.
Automating the database schema evolution process.
Published in The VLDB Journal – The International
Journal on Very Large Data Bases, vol. 22, pp. 73-98.
Genevès, P., Layaïda, N., and Quint, V., 2011. Impact of
XML Schema Evolution. Published in Journal ACM
Transactions on Internet Technology (TOIT), vol. 11,
article 4.
Halevy, A., Rajarama, A., and Ordille, J., 2006. Data
Integration: The Teenage Years. Proceedings of the
32nd International Conference on Very large data
bases, pp 9-16. Seoul, Korea.
Halevy, A., Sarma, A. D., and Dong, X., 2008.
Bootstrapping pay-as-you-go data integration systems.
Proceeding of the 2008 ACM SIGMOD International
Conference of Data, pp. 861-874. Vancouver, Canada.
Joung, Y., and Chuang, F., 2009. OntoZilla: An ontology-
based, semi-structured, and evolutionary peer-to-peer
network for information systems and services. Journal
of Future Generation Computer Systems, vol. 25, n° 1,
pp. 53-63.
Kantere, V., Tsoumakos, D., and Sellis, T. ,2008. A
framework for semantic grouping in P2P databases.
Published in Journal Information Systems, vol. 33, pp.
611-636.
Montanelli, S., Bianchini, D., Aiello, C., Baldoni, R.,
Bolchini, C., Bonomi, S., Castano, S., Catarci, T.,
Antonellis, V., Ferrara, A., Melchiori, M., Quintarelli,
E., Scannapieco, M., Schreiber, A., and Tanca,
L.,2011. The ESTEEM platform: enabling P2P
semantic collaboration through emerging collective
knowledge. Published in Journal of Intelligent
Information Systems, vol. 36, n° 2.
Pires, C. E., Santiago, R., Kedad, Z., Bouzehoub, M. and
Salgado, A. C., 2012. Ontology-based Clustering in a
Peer Data Management System. Published in
International Journal of Distributed Systems and
Technologies (IJDST), vol. 3, Issue 2, pp. 1-21.
Raftopoulou, P., and Petrakis, E. G. M., 2008. A Measure
for Cluster Cohesion in Semantic Overlay Semantic.
Proceedings of the 2008 ACM workshop on Large-
Scale distributed systems for information retrieval.
Napa Valley, USA.
Rijsbergen, C. J., 1979. Information Retrieval, 2
nd
Edition,
MA: Butterworths.
Roth, A., and Skritek, S., 2013. Peer Data Management. In
Data Exchange, Information and Streams, vol. 5, pp.
185-215.
Silva, E. R., Salgado, A. C., 2013. Load Balance for
Semantic Cluster-based Data Integration Systems.
Proceeding of the 17th International Database
Engineering & Applications Symposium (IDEAS’13).
Barcelona, Spain.
Sockut, G. H., and Iyer, B. R, 2011. Online
Reorganization of Databases. Published in Journal
ACM Computing Surveys (CSUR), vol. 41, article 14.
Terwilliger, J. F., Bernstein, P. A., and Unnitha, A., 2010.
“Worry-Free Database Upgrades: Automated Model-
Driven Evolution of Schemas and Complex
Mappings”. Proceedings of the ACM SIGMOD
International Conference on Management of Data, pp.
1191-1194. Indianapolis, USA.
Tian, Y., Song, B., and Huh, E. N.. “Dynamic content-
based cloud data integration system with privacy and
cost concern”. Proceedings of the 8th Annual
Collaboration, Electronic messaging, Anti-Abuse and
Spam Conference, pp. 193-199. Redmond, USA.
2011.
Wall, B., and Angryk, R., 2011. Minimal Data Sets vs.
Synchronized Data Copies in a Schema and Data
Versioning System. Proceedings of the 4th workshop
on Workshop for Ph.D. students in information &
knowledge Management, pp. 67-74. Glasgow, United
Kingdom.
W3C. “OWL – Web Ontology Language”, 2013.
Available in http://www.w3.org/TR/owl-features.
Accessed on October 1
st
..
Zamboulis, L., Martin, N., and Poulovassillis, A., 2010.
Query performance evaluation of an architecture for
fine-grained integration of heterogeneous grid data
sources. Published in Journal Future Generation
Computer Science Systems, vol. 26, pp. 1073-1091.
ICEIS2014-16thInternationalConferenceonEnterpriseInformationSystems
98
