116
for a given instance in
I
a
. Similar to the approach for the instance categorization
problem, we will start with the scores
s(
i, j) to compute an initial ordered list of
tags for instance
i. Likewise, we investigate whether the use of instance relatedness
can lead to improvements over the initial tagging.
When addressing the instance categorization problem, we assumed the relation
between instances and tags to be functional. That is, each instance in
I
a
was as-
sumed to be related to at most one tag (e.g. a
genre or art style). When dealing with
the instance tagging problem however, we assume that multiple tags are applicable
to a given instance. Thus the question is which of the tags are most applicable and
to what extent.
The use of the score
s(
i, j) is a first approximation to identify the tags most
related to the given instance
i. Similar to the computation of the final mapping
m,
we use the similarity between the instances in
I
a
to obtain a final score.
The degree of relatedness of an instance
i
0
to
i is given by
t(
i, i
0
). For tag
j, the
degree of applicability of
j to
i is given by
s(
i, j).
We use the computed scores of relatedness
t(
i, i
0
) to improve the initial tagging
s(
i, j). If two instances are closely related, we expect similar tags for the two.
Hence, if
i
0
is closely related to
i, we want
s(
i
0
, j) to contribute significantly to the
final score
p(
i, j). Using the normalized scoring functions, we can compute the
applicability
p
0
(
i, j) of tag
j to instance
i as follows
p
0
(
i, j) =
∑
i
0
,i
0
6=
i
t(
i, i
0
)
· s(
i
0
, j)
.
(6.6)
If erroneously a high score is found for
s(
i, j), this error is decreased when
close related instances
i
0
have low scores for
s(
i
0
, j).
However,
p
0
(
i, j) does not suffice as no self-relatedness score
t(
i, i) is defined.
We do consider
s(
i, j) relevant when computing the scores for the tags with respect
to instance
i. Hence, we introduce a weight
w for
s(
i, j) as a substitute for
t(
i, i) in
the score
p(
i, j),
Dostları ilə paylaş: