Databases Reference
In-Depth Information
to also handle condition algebra as proposed in
Bohannon et al.
[
2006
]. For example,
assume that matcher A matches
.
R
:
CardInfo
:
cardNum
;
S
:
HotelCardInformation
:
clientNum
;
R
:
CardInfo
:
type
D
'
RoomsRUs
'/
with a certainty of 0:8 and matcher B matches
.
R
:
CardInfo
:
cardNum
;
S
:
HotelCardInformation
:
clientNum
;
Null
/
with a certainty of 0:5. If we use max as the aggregation function, then the entry
.
R
:
CardInfo
:
cardNum
;
S
:
HotelCardInformation
:
clientNum
/
in the combined similarity matrix will contain the tuple
h
0:8;
R
:
CardInfo
:
type
D
'
RoomsRUs
'
.
Both context conditions and the constraint function serve in constraining the
possible space of correct matchings. Traditionally, wasassumedtobegivenby
the designer (e.g., a 1
W
1 matching constraint or a user feedback after an iteration
of the matching process). Context, on the other hand, is learned from the domain.
However, there is no dichotomy here. One can assume that certain context condi-
tions are given by the designer while some matching constraints are learned from
the application. A decision on whether a constraint should be specified in the simi-
larity matrix or as part of the constraint function can be based on the following two
observations:
i
1.
The similarity matrix captures information that is inherently uncertain while
is defined to be deterministic. Therefore, a context provided by a designer can
be considered deterministic and become part of while learned contexts can be
embedded at the matrix level.
2.
The similarity matrix inherently captures attribute correspondence information
while can handle schema level constraints. For example, using only the
similarity matrix, one can assume that the contextual attribute correspondence
.
R
:
CardInfo
:
cardNum
;
S
:
HotelCardInformation
:
clientNum
;
R
:
CardInfo
:
type
D
'
RoomsRUs
'/
can coexist with the attribute correspondence
.
R
:
CardInfo
:
cardNum
;
S
:
CardInformation
:
cardNum
/
simply because both entries have a nonzero similarity measure. It is only at the
constraint function level that such coexistence can be explicitly forbidden.
Therefore, schema level contextual constraint should be modeled using while
attribute level constraints are modeled as part of the similarity matrix.
