Multi-domain Schema Clustering and Hierarchical Mediated Schema
Generation
Qizhen Huang, Chaoliang Zhong and Jun Zhang
Information Technology Lab, Fujitsu R&D Center Co., LTD.,
No.56 Dong Si Huan Zhong Rd, Chaoyang District, Beijing, P.R. China
Keywords: Data Integration, Hierarchical Mediated Schema, Schema Clustering.
Abstract: In data integration, users can access multiple data sources through a uniform interface. Yet it is still not easy
to query from data sources where many domains coexist even if the data sources are clustered into several
domains since users have to write different query clauses for each different domain. Previous researches
have presented various data integration techniques, but nearly all of them require the schemas of data
sources to be integrated belong to the same domain, or failed to address that some different domains may be
the sub-domains of a high level domain in which case a more abstract query clause for upper domain can
substitute several less abstract query clauses for lower domains. In this paper, we propose a graph-based
approach for clustering schemas which would finally expose to users a hierarchical mediated schema forest,
and a query forwarding mechanism to transform queries down along the schema forest. A set of
experimental results demonstrate that our schema clustering algorithm is effective in clustering the data
sources into hierarchical schemas, queries on the mediated schemas could achieve answers with good
accuracy, and the cost of writing query clauses for users is reduced without losing query accuracy.
1 INTRODUCTION
Many application contexts involve a large amount of
heterogeneous structural data sources, such as
HTML tables, downloadable spreadsheets and tables
from enterprise databases. If users need to obtain
answers from the underlying data sources that meet
certain conditions, it’s troublesome to access each
data source one by one. Data integration is an
approach used to access such large numbers of
heterogeneous structured data sources by providing
users with a uniform interface to these data sources
(Sarma et al., 2008; Dong et al., 2007; Aboulnaga
and Gebaly, 2007). A data integration system
consists typically of a mediated schema for one
domain and semantic mappings between the schemas
of the data sources and the mediated schema (Sarma
et al., 2008). With data integration, the user can pose
queries on the mediated schema so as not to pose
queries on each table of different data sources to get
the answers. It is now common that data sources
where many domains coexist are maintained by one
system. Yet it is not easy to query from such system
even if the data sources are clustered into several
domains and data integration are applied for each
domain since users still have to choose the target
domain or write different query clauses for different
domains when querying.
Previous researches present various data
integration techniques, but nearly all of them require
the schemas of different data sources to be integrated
belong to the same domain (Sarma et al., 2008;
Dong et al., 2007; Aboulnaga and Gebaly, 2007).
When dealing with schemas of multiple domains,
data integration without the step of clustering
schemas tends to produce semantically incoherent
mediated schemas and incorrect schema-mappings.
Clustering schemas has been taken into
consideration in (Mahmoud and Aboulnaga, 2010).
Their method clustered real schemas of databases
into different domains, but in this case it is difficult
for users to identify which mediated schema is their
target schema because users still need to face too
many mediated schemas. This observation led to the
requirement of an obvious mark to represent a
mediated schema and to reduce the amount of
mediated schemas that user would visit so that the
cost of accessing underlying data can be diminished.
This paper proposes an approach to generate
hierarchical mediated schemas for different domains.
111
Huang Q., Zhong C. and Zhang J..
Multi-domain Schema Clustering and Hierarchical Mediated Schema Generation.
DOI: 10.5220/0004825601110118
In Proceedings of the 16th International Conference on Enterprise Information Systems (ICEIS-2014), pages 111-118
ISBN: 978-989-758-027-7
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
Specifically, this paper (1) newly defines a similarity
measurement between two schemas considering the
different importance of schema name and attributes
simultaneously, (2) designs a graph-based algorithm
to hierarchically cluster schemas through two phase
removals of edges and generate mediated schemas
for each schema clusters adding mediated schema
names, (3) and designs a query forwarding
mechanism to transform queries down along the
hierarchical mediated schemas. Mediated schema
name is first proposed in this paper to represent a
mediated schema which is helpful for the user to
identify her target schemas. Hierarchical schemas
are used to reduce the amount of mediated schemas
which users have to visit. Finally, we design the
query forwarding mechanism to conform to the
hierarchical schemas framework.
The remainder of the paper is organized as
follows. Section 2 presents the related work to our
research. Section 3 gives an overview of our
solution. In Section 4, we show the details of
hierarchical mediated schema generation. Section 5
explains how the queries are executed in the
database with hierarchical mediated schemas.
Experimental results and analysis are presented in
Section 6. Finally, we draw conclusions in Section 7.
2 RELATED WORK
Most of the previous work on structural data
integration focused on automatically creating
mediated schemas (Sarma et al., 2008) and schema-
mapping creation (Sarma et al., 2008; Dong et al.,
2007). Our approach automatically groups the
schemas of different domains while traditional one-
domain schema generation methods chose the
schemas belonging to one domain to integrate.
Previous work on schema-mapping creation (Sarma
et al., 2008; Dong et al., 2007) mainly studied on
how to compute a set of attribute correspondences
between attributes in mediated schemas and
attributes in source schemas. We create all the
schema-mappings by assign a unique schema ID to
each schema to avoid the computation of attribute
correspondences. In (He et al., 2004), the authors
take clustering schemas into domains into
consideration. However, their approach assumes that
for each domain there are some anchor attributes that
do not appear except in that domain, but we do not
require such prerequisite. Schema clustering is also
mentioned in (Madhavan et al., 2007) as part of their
pay-as-you-go architecture, but they did not describe
the details such as similarity between schemas and
the clustering process. The work closest to ours is in
(Mahmoud and Aboulnaga, 2010) which employed
hierarchical clustering algorithm (Murphy, 2012) to
cluster schemas into domains and assigned schemas
to domains using a probabilistic model. Compared
with the multi-domain schema clustering method in
(Mahmoud and Aboulnaga, 2010), the name of
schema and attributes are all considered and given
different importance in our work, while their method
only considers the attributes when computing the
similarity of two schemas. Furthermore, our
approach produces a hierarchical mediated schema
forest with the lowest level schemas being the real
source schemas but other level schemas being the
abstract mediated schemas. We produce mediated
schema both for source schemas and mediated
schemas, whereas, it (Mahmoud and Aboulnaga,
2010) only produces source schema clusters
(mediated schemas of one level). The final
difference between our approach and other data
integration methods is that we add a name to
mediated schema to help users understand a schema
much more.
3 SOLUTION OVERVIEW
The objective of our research is to automate
producing hierarchical mediated schemas for data
sources of multiple domains. Providing a valid and
complete representation of a domain is a non-trivial
task in Conceptual modeling (Wand and Weber,
2002). Here we follow the definition of domain in
(Mahmoud and Aboulnaga, 2010): a domain is a set
of schemas with sufficiently large intra-domain
similarity and sufficiently large inter-domain
dissimilarity, according to some measure of
similarity. A schema is defined as a set of attributes
associated with the name of this schema in our paper.
A database table with its name and without its data is
such a schema accordingly.
The input of our system is a set of schemas that
are extracted from structured data sources. The
attributes and the schema names are the only
information we require. Domains and their upper
domains would be discovered from the available
schemas. Then the discovered domains would be
delivered to a schema mediation algorithm which
has been widely studied. Consequently, a mediated
schema forest which is the final output of our system
will be exposed to users who want to query the
underlying data sources. Schema forest in this paper
is defined as hierarchical schemas where schemas in
the lowest level are source schemas and other lower
ICEIS2014-16thInternationalConferenceonEnterpriseInformationSystems
112
level mediated schemas are sub-domains of their
respective high level domains.
Figure 1: A scenario about domain and sub-domain of
several data sources.
Figure 1 depicts the hierarchical relationship
about schemas in different data sources. As we can
see on the 2-nd level of Figure 1, the domain Patent
contains three source schemas from different sources,
i.e. T
11
of Source 1, T
22
of Source 2 and T
n3
of
Source n. A mediated schema with attributes and
Patent as its schema name represents domain Patent
will be generated. On the 1-st level, the domain
Academic literature owns two sub-domains, i.e.
Patent and Paper. The mediated schemas reflecting
all the domains are generated the same as the domain
Patent.
Figure 2 shows the system diagram of the
proposed approach. In traditional small-scale
database scenario, one user only needs to pose a
query on a table and get the answers she wants.
Steps 1, 4 and 5 of Figure 2 are required in such
case. When dealing with large-scale data space, step
2 is required to offer a uniform interface of all data
sources and step 3 is used to give appropriate query
formats conformed to mediated schemas. We will
introduce the details of step 2 in Section 4 and step 3
in Section 5.
4 MEDIATED SCHEMA FOREST
GENERATION
4.1 Schema Similarity
Before clustering the schemas, we need to determine
how to measure the similarity between two schemas.
Besides the attributes of a schema which are useful
to represent a schema, the schema name is also
valuable since different databases are likely to use
similar and even the same titles to name the tables
that may relate to the same object. For instance, a
table in database
1
is named staff, while another table
in database
2
called employee probably describes the
same object. In such case, these two tables are likely
to belong to the same domain. Thus, we define a new
similarity measurement between two tables/schemas
considering the different importance of schema name
and attributes simultaneously. The definition is as
equation (1).
similaritys
,s
θ∙sims
.name,s
.name
1θ
∑∑
,
∈
.
∈
.
|
.
|
∙|
.|
θ∈[0,1]. (1)
Figure 2: A query system with hierarchical mediated schemas.
Multi-domainSchemaClusteringandHierarchicalMediatedSchemaGeneration
113
The first term after “=” measures the syntactic or
semantic similarity of two schema names. Similarity
between two attribute sets of the respective schemas
is evaluated in the second term. θ is a parameter
which is used to estimate the importance of schema
name. It is pre-defined and could be tuned in
concrete application. s.name represents name of
schema s. The capital letter “A” denotes the attribute
set of a schema. Consequently, |s.A| represents the
cardinality of attribute set A of schema s. sim(a,b)
could be syntactic or semantic similarity of a and b.
4.2 Hierarchical Generation of
Mediated Schemas
An algorithm of hierarchical generation of mediated
schemas is proposed. The details are depicted in
Algorithm 1.
A
lgorithm 1: Hierarchical mediated schema
s
generation.
0: Input: a set of source schemas S
all
1: Output: a set of mediated schemas T
2: Initialize α, , S= S
all
, T=null
3: for each schema s
i
∈S
4: for each schema s
j
∈S, s
i
≠ s
j
5: Compute similarity(s
i
, s
j
) using equation (1)
6: if one of the similarities is greater than
7: Construct a weighted graph G(V,E,W),
where V=S, E={(s
i
, s
j
)}, W={(v
i
, v
j
)| (v
i
, v
j
)=
similarity(s
i
, s
j
) }
8: for each v
i
∈V
9: for each v
j
∈V, v
i
≠v
j
10: if ((v
i
, v
j
) <(v
i
, v
imax
)
where (v
i
, v
imax
) =max (v
i
, v
k
) (v
i
≠v
k
,
v
k
∈V ))
&& ((v
i
, v
j
) <(v
j
, v
jmax
)
where (v
j
, v
jmax
) =max (v
j
, v
k
) (v
j
≠v
k
,
v
k
∈
V))
11: remove (v
i
, v
j
)
12: for each connected component c
i
⊆ G
13: for each e
i
, e
j
∈c
i
.E, ∈c
i
.W
14: if
e
i
/e
j
>= α remove e
j
15: if
e
i
/e
j
<= 1/ α remove e
i
16: S=null
17: for each c
i
⊆G
(where {s
i1
, s
i2
, …}=c
i
.V )
18: create a mediated schema med-s
i
for c
i
19: infer med-s
i
.name by {s
i1
.name,
s
i2
.name, …}
20: add med-s
i
to S
21: med-s
i
.child_set={ s
i1
, s
i2
, …}
22: add med-s
i
to T
23: update α
24: goto step 3
24: else return T
Steps 3-5 in Algorithm 1 compute the pair-wise
schema similarity according to our newly defined
similarity measurement. Step 7 constructs a
weighted graph with schemas as vertices. If the
similarity of two schemas is greater than 0, there is
an edge (s
i
, s
j
) with weight similarity(s
i
, s
j
). Steps 8-
11 are the first phase removal of edges so as to hold
the maximal similarities. The second phase removal
of edges based on larger similarities is taken in steps
12-15. Remove edges in a connected component if
the weight of the edge is significantly less than other
edges’ weights. α is a threshold used to control
whether remove an edge or not in a connected
component. If the orders of magnitude of two edges’
weights are larger than α or less than 1/α, remove the
less weighted edge. α is updated for each level, the
higher lever owns a lager α since the schemas in the
same cluster of higher level are more dissimilar than
lower level. α can be tuned for different schema sets.
In steps 17-22, we can use any mediated schema
generation techniques (Mahmoud and Aboulnaga,
2010) to create mediated schemas since the schemas
belong to one domain for a schema cluster after the
schema clustering process. In this paper, the
mediated schema of each domain is created as the
method in (Sarma et al., 2008). We choose the most
frequent attribute in one cluster as the mediated
attribute of its mediated schema. The mediated
attribute along with the attributes in its cluster are
stored as a schema attribute mapping which will be
described in Section 5.
Besides the mediated schema attributes, unlike
other methods, we provide each schema a name as
an obvious mark to a schema. Each mediated schema
will be given a name according to the names of
including schemas. Cluster labelling method using
Wikipedia (Carmel et al., 2009) is used to infer the
name of mediated schema. A search index was built
from one of the available Wikipedia dump
1
using
Lucene
2
search system. For the names of each
including schemas, we search the generated index
with the disjunction of such names. The titles and
categories of the documents returned by the search
are considered as the potential candidates of the
target mediated schema name. Finally, the most
frequently appeared term of the candidates as well as
all including schema names is chose as the mediated
schema name.
The iteration ends when all of the similarities are
less than the threshold . Each schema graph
corresponds to one level in the schema forest and
1
http://dumps.wikimedia.org/enwiki/20130204/
2
http://lucene.apache.org/
ICEIS2014-16thInternationalConferenceonEnterpriseInformationSystems
114
can produce a set of mediated schemas. The above
procedures correspond to step 2 in Figure 2.
5 QUERY FORWARDING
At the schema clustering time, we produce a
mediated schema forest of various data sources. At
query phase, users can pose queries using the
terminology of any of the mediated schemas
although most of time the higher level schemas are
used. The query is reformulated and delivered
downward to the source schemas according to the
mapping between the mediated schema and its lower
schemas. A whole mapping called schema mapping
in our system is designed to consist of two parts.
First part is the one-to-many schema name mapping.
The mediated schema name is mapped downward to
the schema names of the lower schemas. We give an
ID to each schema which can uniquely identify the
schema. The name mappings keep schema names
along with their IDs. The IDs guarantee the schema
mappings are correct mappings. Consider the next
example of 3 source schemas.
Example 1:
(Paper: Paper_No., Title, Presentation,
Author_info, Nation, Contact, Company)
(Proceedings: Year, Conference, Title, Author1,
Company1, Author2, Company2)
(Publication: Ref_No., Type, Title, Author_list,
Reference, Volume, Issue_section, Pages, Year,
Month, Day, Conference_notes)
In the above example, schema name and schema
attributes are separated by colon, and attributes are
separated by comma. After processing by our system,
we could obtain the following mediated schema.
Example 2:
(Academic_publication: Title, No., Author,
Conference, Year)
The name mapping between the above mediated
schema and source schemas is as follow.
Example 3:
((4, Academic_publication) → ((1, Paper), (2,
Proceedings), (3, Publication)))
This mapping maps schema (4,
Academic_publication) to three lower schemas (1,
Paper), (2, Proceedings), and (3, Publication). The
first element in (4, Academic_publication) is the
schema ID and the other one is the schema name.
One-to-many attribute mappings are the other
part of the schema mapping. The attribute mapping
format is the same as the name mapping, but the
second element within the parenthesis is an attribute
such like “Title” in (4, Title) of Example 4. There
would be several attribute mappings in one schema
mapping since a mediated schema usually owns
several mediated attributes. Consider the next
mappings.
Example 4:
((4, Title) → ((
1, Title), (2, Title), (3, Title)))
((4, No.) → ((1, Paper_No.), (3, Ref_No.)))
((4, Author) → ((1, Author_info), (2, Author1), (2,
Author2), (3, Author_list)))
((4, Conference) → ((2, Conference), (3,
Conference_notes)))
((4, Year) → ((2, Year), (3, Year)))
Example 4 is the attribute mappings between
Example 1 and Example 2.
When users pose a query on the mediated schema
in Example 2, the system will reformulate the query
to the source schema according to the mappings in
Example 3 and Example 4. Several queries would be
generated matching the source schemas in Example
1. Finally the databases will return the required
answers to users. To illustrate the forwarding process,
consider the next queries. The query is posed on the
mediated schema in Example 2.
SELECT Title, Author, Year
FROM Academic_publication
WHERE Year
>
2006 AND Author = ‘Strehl’
According to the mappings in Example 3 and 4,
the next two queries will be posed on the
corresponding data sources automatically.
(1) SELECT Year, Title, Author1, Author2
FROM Proceedings
WHERE Year
>
2006 AND (Author1= ‘Strehl’
OR Author2 =‘Strehl’)
(2) SELECT Title, Author_list, Year
FROM Publication
WHERE Year
>
2006 AND Author_list = ‘Strehl’
6 EXPERIMENTS
Our algorithms were implemented in Java. We run
the experiments on a Windows 7 machine, with
2.60GH Intel(R) i5 processor and 8GB memory. The
goal of our experiments is to demonstrate that our
schema clustering algorithm is effective in clustering
the data sources of multiple domains, queries on the
mediated schemas could achieve answers with good
accuracy and the cost of writing query clauses for
users is reduced without losing query accuracy.
For the purpose of our query evaluation, we used
MySQL to store the data. Two string similarity
measurements are utilized to compute the schema
similarity since two strings may be semantically
Multi-domainSchemaClusteringandHierarchicalMediatedSchemaGeneration
115
similar or syntactically similar. For example, “Film”
and “Movie” are semantically similar, whereas
“Author” and “Author_info” would be syntactically
similar. We use the SecondString tool
3
and WS4J
4
to compute the syntactic similarity and semantic
similarity of pair-wise attribute strings or schema
names. The function sim() in equation (1) is
evaluated by the larger one of syntactic similarity
and semantic similarity.
6.1 Data Used
Table 1: Number of schemas in each domain and each
upper domain.
#schema
(the leaf
layer)
#schema
(the third
layer)
#schema
(the
second
layer)
#schema
(the first
layer)
Paper
23
2 1 1
Patent
16
Employee
26
2 1 1
Student
15
Car
12
3 1 1
Truck
16
Bus
13
Movie
25
3 1 1
TV serial
11
Novel
17
Baseball
15
2
3 1
Basketball
17
Road
running
13
2
Adventure
running
10
Swing
16
2
Diving
12
We collect the schemas and their data from our
enterprise databases and the downloadable
spreadsheets that are obtained using Google’s
“search by file type” option with certain keywords
related to different domains. For the purpose of
generating hierarchical mediated schema forest,
some domains we selected expose parallel
relationship and we could infer the upper domains
for them. For instance, “Employee” and “Student”
belong to the upper domain “Person”. In each
spreadsheet, we use the strings in the headers of the
columns as the attributes, and manually choose the
title or topic of the spreadsheet as the schema name.
For the tables extracted from our enterprise
databases, the table names and their field names are
used as schema names and attributes. We manually
3
http://secondstring.sourceforge.net/
4
https://code.google.com/p/ws4j/
associated each source schema a label indicating the
domain it belongs to. And schema name of a
mediated schema indicates the domain it represents.
Table1 shows some details of the extracted tables.
6.2 Schema Clustering Evaluation
In this part, we evaluate the effectiveness of the
proposed schema clustering algorithm. Let the set S
={s
1
, s
2
, …} denote the schema set. d={d
1
, d
2
, …}
denotes the actual domain set associated to each
source schema. Two functions indicating which
domain a schema is contained in are defined as
D
before
(s): { s
1
, s
2
, …}→{ d
1
, d
2
, …},
D
after
(s): { s
1
, s
2
, …}→{ d
1
, d
2
, …}.
Specifically, D
before
(s) = d
i
(d
∈d) means the
schema s are associated with domain d
i
, and D
after
(s)
= d
i
means s are clustered into the cluster with d
i
as
its domain, i.e. the mediated schema name of the
cluster.
We measure precision, recall and NMI
(Normalized Mutual Information) (Strehl and Ghosh,
2003) of clustering results of the leaf layer schemas
in the resulting mediated schema forest since the
domain of each mediated schema cannot be assigned
in advance and the effectiveness of schema
clustering algorithm can be validated using the
schemas in any layer.
Precision: We estimate the average precision as
1
|d|
|TP
|
|TP
||FP
|
∈
where TP
is the true positive set of domain d
and
FP
is the false positive set of domain d
. Therefore,
TP
={s| D
before
(s) = d
i
&& D
after
(s) = d
i
}, and
FP
={s| D
before
(s) d
i
&& D
after
(s) = d
i
}.
Recall: We estimate the average recall as
1
|d|
|TP
|
|TP
||FN
|
∈
where FN
is the false negative set of domain d
.
Therefore FN
={s| D
before
(s) = d
i
&& D
after
(s) d
i
}.
NMI: We estimate the NMI according to the
definition in (Strehl and Ghosh, 2003).
In this paper, we emphasise that the schema
names are important in clustering schemas. To
evaluate such view, we run the experiments with
different values of θ on three schema sets composed
of the schemas showed in Table 1 to see how the
clustering results change. We selected the schemas
of different domains the schema names of which are
clearly not similar to constitute schema set SS1, but
in schema set SS2, some schema names in different
ICEIS2014-16thInternationalConferenceonEnterpriseInformationSystems
116
domains are somewhat similar to others. Schema set
ALL is comprised of all the schemas in Table 1. The
clustering results are showed in Figure 3. θ=0 means
the schema names are not concerned in similarity
measurement when clustering schemas. As θ
increases, the three metrics increase, but they would
not increase all the time. Each measurement reaches
a maximal value for different schema sets, and then
decreases. The parameter α influences the clustering
results as well. If α is too small, there would be more
small clusters or some schemas are isolated alone. In
contrary, a too large α will lead less clusters, namely,
the schemas are not sufficiently separated. Before
evaluating the influence of the primary factor θ, we
conducted the experiments multi-times to find an
appropriate α for each schema set. The results in
Figure 3 prove that schema names are useful in
estimating the domain a schema belongs to, but one
schema name could not represent a schema.
With appropriate θ and α for each schema set, we
run experiments to compare the schema clustering
results of hierarchical clustering and our algorithm.
We implemented the hierarchical clustering
algorithm proposed in (Mahmoud and Aboulnaga,
2010). The comparison results are showed in Table
2.
Table 2: Comparison of clustering results.
SS1 SS2 ALL
Hierarchical
clustering
Precision 0.81 0.80 0.80
Recall 0.91 0.89 0.89
NMI 0.77 0.75 0.76
Our
algorithm
Precision 0.82 0.78 0.81
Recall 0.94 0.87 0.89
NMI 0.79 0.73 0.77
For the three different schema sets, the three metrics
of SS
1
are higher than that of SS2, since SS2
contains some ambiguous schemas. It is not
surprising that the accuracy of schema set ALL is
between that of SS1 and SS2. Through Table 2, we
can also see that the three metrics of our algorithm
are a bit lower than hierarchical clustering for SS2. It
is because that the hierarchical clustering using
single linkage is good at clustering ambiguous
schemas into isolated clusters. However, the three
metrics are higher than hierarchical clustering for
SS1 and ALL. To sum up, our algorithm is effective
in clustering schemas.
6.3 Query Quality
In this part, we evaluate the quality of queries posed
on the mediated schema forest generated by our
method so that we can prove the practicability of
hierarchical meditated schemas. For each domain in
each layer of the schema forest, we created 8 diverse
queries. The queries are created by the following
principles. Each query contains one to three
attributes in the SELECT clause and zero to three
predicates in the WHERE clause. The table names in
the FROM clause are the mediated schema names.
The attributes in the SELECT and WHERE clauses
are attributes from the exposed mediated schemas.
Figure 3: Influence of Schema Name. the Value of Theta
(Θ) Indicates the Importance of Schema Name.
Multi-domainSchemaClusteringandHierarchicalMediatedSchemaGeneration
117
In our experiments we used two metrics
precision and recall to measure the query accuracy.
The two metrics are different with that used in
clustering evaluation. For all queries, let A be the set
of answers generated by our approach and B be the
set of true answers. We estimate query precision as
|⋂|
||
and recall as
|⋂|
||
. We compare the average
query accuracy of our algorithms with that of one
mediated schema method and key word query
method. The results are showed in Table 3.
Table 3: Comparison of query quality.
SS1 SS2 ALL
One mediated
schema
Precision 0.87 0.81 0.83
Recall 0.78 0.72 0.76
keyword
Precision 0.85 0.75 0.82
Recall 0.74 0.68 0.72
Our algorithm
Precision 0.97 0.88 0.86
Recall 0.88 0.84 0.84
Table 3 states obvious that the query precision
and recall of our algorithm gain a notable increase
than that of another two methods. Not surprisingly,
keyword query method has the worst results since
keyword search engines of database offer a simple
and general solution for searching any kind of
information. One mediated schema method treats all
schemas as schemas in one domain. That is the
reason why its results are worse than our algorithms.
We have observations that the precision of our
method is high. It is because the schema mappings
we designed are correct mappings between mediated
schemas and source schemas due to the schema ID.
Generally speaking, the results demonstrate that
queries on the mediated schemas generated by our
algorithm could achieve answers with good accuracy.
Finally, we compare the average query accuracy
of one level mediated schemas and hierarchical
mediated schemas. For the source schemas of Sports
in the bottom of Table 1, we need to write 6 queries
in the domain of the source schema clusters, but only
one query clause in the highest level domain. These
two situations obtain exactly the same answers. The
similar behaviours of less query clauses with the
same answer are presented for the other source
schemas. This is due to the query forward
mechanism which transforms the query from the
highest level domain to the lowest level domain
transparently for users. In one word, the cost of
writing query clauses for users is reduced without
losing query accuracy.
7 CONCLUSIONS
We believe that users need hierarchical mediated
schemas with names when facing a large number of
data sources from different domains. Therefore, in
this paper, we presented an approach of generating
hierarchical mediated schema forest with schema
name assigned to each mediated schema for data
sources of multiple domains. We also explained the
compatible query execution process against the
hierarchical mediated schemas. The experimental
results on data sources of different domains validate
the feasibility and effectiveness of our approach.
Users can easily find the target mediated schema and
obtain the answers with good accuracy and less
query effort.
In future, we will investigate more on when we
should stop the process of producing hierarchical
mediated schema to reduce the time cost and applied
the approach to very large scale source schemas.
REFERENCES
Sarma, D.A., Dong, X. and Halevy, A., 2008.
Bootstrapping pay-as-you-go data integration systems.
In SIGMOD, ACM.
Dong, X., Halevy, A. and Yu, C., 2007. Data integration
with uncertainty. In VLDB, ACM.
Aboulnaga A. and Gebaly, K.E., 2007. μbe: User guided
source selection and schema mediation for internet
scale data integration. In ICDE, IEEE.
Mahmoud, H.A. and Aboulnaga, A., 2010. Schema
clustering and retrieval for multi-domain pay-as-you-
go data integration systems, In SIGMOD, ACM.
He, B., Tao, T. and Chang, K. C.-C., 2004. Organizing
structured web sources by query schemas: a clustering
approach. In CIKM, ACM.
Madhavan, J., Cohen, S., Dong, X., Halevy, A., Jeffery, S.,
Ko, D. and Yu, C., 2007. Web-scale data integration:
you can afford to pay as you go. In CIDR.
Murphy, K., 2012. Machine learning: a probabilistic
perspective. MIT Press, London.
Wand, Y. and Weber, R., 2002. Research Commentary:
Information Systems and Conceptual Modeling--A
Research Agenda. Information Systems Research,
vol.13(4), pp.363-376.
Carmel, D., Roitman, H. and Zwerdling, N., 2009.
Enhancing Cluster Labeling Using Wikipedia. In
SIGIR, ACM.
Strehl, A. and Ghosh, J, 2003. Cluster ensembles: a
knowledge reuse framework for combining multiple
partitions, The Journal of Machine Learning Research,
vol.3, pp.583-617.
ICEIS2014-16thInternationalConferenceonEnterpriseInformationSystems
118
