Databases Reference
In-Depth Information
A simple example of aggregation function is demonstrated with BMatch
[
Duchateau et al. 2008b
] or Cupid [
Madhavan et al. 2001
]. Their authors aggre-
gate the results of terminological measure with the ones computed by a structural
measure by varying the weights applied to each measure (
2
and
2
,
3
and
3
,etc.).
In most tools, default values are given to these weights. They are mainly the
results of intensive experiments. For example, the default weights of COMA
's
name
and
data type
similarity measures are 0:7 and 0:3, respectively [
Do and Rahm
2002
]. As explained in Glue [
Doan et al. 2003
] or APFEL [
Ehrig et al. 2005
], it
is possible to tune the weights of an aggregation function automatically, thanks to
machine learning techniques.
To help tuning the weights in aggregating functions, we discuss the iMAP
approach [
Dhamankar et al. 2004
]. This matcher mainly provides a new set of
machine learning-based measures for discovering specific types of complex map-
pings (e.g.,
name
is a concatenation of
firstname
and
lastname
). It also includes an
explanation module to clarify why a given correspondence has been discovered to
the detriment of another candidate. For instance, this module is able to describe that
a string-matching classifier has a strong influence for a discovered correspondence.
Thus, user can use this feedback to decrease the weight of this classifier.
CC
5.3
Supporting Users to Revise Strategies
Although most matchers simply provide a graphical user interface to visualize the
results, recent works have pointed out a need for selecting the best strategy. For
instance, including some mechanisms to easily update the weights of a function so
that users can directly analyse impacts of these changes.
Here, we describe recent works that aim at supporting users during the tasks of
selecting appropriate similarity measures and combining them. To combine them
efficiently, weights have to be efficiently tuned. To support users during these tasks,
two tools have been designed: AgreementMaker and Harmony. Whatever the tech-
nique they use (interactions with users or strategy filters), they enable a revision of
the current strategy by adding, removing or modifying parameters and similarity
measures involved in the combination. We further describe each of these tools in the
rest of this part.
5.3.1
AgreementMaker
The originality of AgreementMaker [
Cruz et al. 2007
,
2009
] is the capability of
matching methods combination. Moreover, it provides facilities for tuning manually
the quality of matches. Indeed, one of the interesting features of AgreementMaker is
a comprehensive user interface supporting both advanced visualization techniques
and a control panel that drives the matching methods. This interface, depicted by
Fig.
10.5
, provides the user facilities to evaluate the matching process, thus enabling
the user to be directly involved in the loop and evaluation strategies.
