Information Technology Reference
In-Depth Information
Table 2.2
The statistics of
NUS-WIDE-USER15
Users
|
U
|
Images
|
I
|
Tags
|
T
|
|
O
|
USER15
3,372
124,099
5,018
1,223,254
owning no less than 15 images and keep their images to obtain our experimental
dataset, which is referred as NUS-WIDE-USER15. Table
2.2
summarizes the col-
lected dataset.
|O|
is the number of observed triplets. The NUS-WIDE provides
ground-truth for 81 tags of the images. In the experiments, we evaluate the perfor-
mance of tag refinement by the F-score metric:
2
×
Precision
×
Recall
Fscore
=
(2.22)
Precision
+
Recall
2.4.2 Parameter Settings
The proposed approach, RMTF, has five parameters, the rank of factor matrices
r
U
,
r
I
,
r
T
and the regularization weights
. We explore the influence of different
parameter settings on a smaller but representative dataset, NUS-WIDE-USER50,
which has 588 users and 55,141 images by filtering out the users with fewer than 50
images.
Choosing the rank of factormatrices
r
U
,
r
I
, and
r
T
inTucker Decompositionmodel
is not trivial. A practical option is to use ranks indicated by SVD on the unfolded
matrices in each mode [
1
]. The tensor
ʱ
,
ʲ
can be unfolded along different modes,
leading to three newmatrices
Y
U
∈ R
|U|×|I||T|
,
Y
I
∈ R
|I|×|U||T|
and
Y
T
∈ R
|T|×|U||I|
.
In this way,
r
U
,
r
I
, and
r
T
are chosen by preserving a certain percentage of singular
values in the unfolded matrices. By fixing small values of
Y
001,
we investigated the average F-score of tag refinement on NUS-WIDE-USER50 by
tuning the percentage of the preserved energy from 50 to 95%. The result in Fig.
2.4
a
ʱ
=
0
.
001 and
ʲ
=
0
.
Fig. 2.4
Impact of parameters (a) rank numbers (b)
ʱ
and
ʲ
. ©[2012] IEEE. Reprinted, with
permission, from Ref. [
34
]
