Information Technology Reference
In-Depth Information
Fig. 2.2
Tagging data interpretation.
a
0/1 scheme,
b
ranking scheme. ©[2012] IEEE. Reprinted,
with permission, from Ref. [
34
]
K
X
IT
i
Top
(
i
,
K
)
=
max
t
(2.6)
:
∈T
In the experiment, we fix
K
=
10.
2.3.1 Ranking-Based Optimization Scheme
Traditional factorization models [
19
,
47
] approximate the tagging data based on
the
0/1 scheme
. Under the situation of social image tagging data, the semantics of
encoding all the unobserved data as 0 are incorrect, which is illustrated with the
running example in Fig.
2.2
a:
•
First, the fact that
user3
has not given any tag to
image2
and
image4
does not mean
that
user3
considered all the tags are bad for describing the images.
4
Maybe he or
she does not want to annotate the image or has no chance to see the image.
•
Secondly,
user1
annotates
image1
with
tag3
only. It is also unreasonable to assume
that other tags should not be annotated to the image, as some concepts may be
missing in the user-generated tags and individual user may not be familiar to all
the relevant tags in the large tag set.
According to the optimization function in Eq. (
2.3
), the learning process tries to
predict 0 for both cases, which is apparently unreasonable. To address the above
4
We call triplets like
(
u
3
,
i
2
,
:
)
and
(
u
3
,
i
4
,
:
)
as the neutral triplets.