Databases Reference
In-Depth Information
8
Discussion
These diverse areas differ greatly in terms of goal, context, data model and more,
but there are a number of similarities. We now discuss some of the similarities
and differences across these different research areas. Because evaluating the qual-
ity of automatic merging algorithms is an open problem, we do not compare the
approaches based on quality.
8.1
Separation of Match and Merge
Ontology merging, view integration, and CSCW all separate matching and merg-
ing. Ontology merging calls the difference alignment and merging [
Noy and Musen
2000
], but despite defining them as separate problems, the techniques that they use
still force the system to perform both actions at the same time. The work on view
integration also requires a matching step to occur before the merging process can
begin. However, as it processes the matched elements, it may discover more matches
that can be made to help with the merging process [
Spaccapietra and Parent 1994
].
CSCW and Model Management make a complete separation between matching and
merging. A question is thus how much less efficient does it become to completely
divorce the merging from the matching. In some cases, interleaving the matching
with the merging (e.g., as in view integration) can cut down on the initial matching
that needs to be done. However, there has been substantial work on schema match-
ing as an independent problem (see [
Rahm and Bernstein 2001
;
Doan and Halevy
2004
] for some surveys). This increases the likelihood that future works on merging
schemas will use the results of these matching algorithms as input and thus schema
merging and matching will become more, rather than less, distinct over time.
The work in
Radwan et al.
[
2009
] represents a pull in the other direction, as it
exploits information about the potential matches to suggest merge results. It would
be interesting to see how the work in
Radwan et al.
[
2009
] can extend the work of
various schema matchers.
8.2
Treating Models Asymmetrically
One idea that occurs repeatedly in all of these works is that the models are treated
asymmetrically, allowing for the algorithms to function more automatically in the
presence of conflicting choices. This may be because one model is more general or
stable than the other and thus assumed to be the “preferred” model [
Noy and Musen
1999a
]. This allows merging operations to proceed much more automatically by
giving it a clear indicator which element to prefer in the case of a conflict.
