Information Technology Reference
In-Depth Information
x
10
4
14
12
10
8
6
4
5
10
15
20
25
30
35
40
Iterative Number
Fig. 2.3
The convergence curve of the learning algorithm. ©[2012] IEEE. Reprinted, with permis-
sion, from Ref. [
34
]
learnt factors, tag refinement is performed by computing the cross-space
image
-
tag
associations.
In the experiments, we observed that the proposedALAconverges to theminimum
after about 20 iterations. Figure
2.3
shows the change of objective function values
in the convergence process. We perform our experiments on MATLAB in a PC with
2.13GHzCPUand 16GBmemory. The convergence time on the experimental dataset
is about 6 hours. Actually, in the proposed learning algorithm, each factor vector
i
m
is updated independently of other vectors, which gives rise to potentially massive
parallelization (e.g. parallel MATLAB). Theoretically, the algorithm achieves a lin-
ear converge speedup which is proportion to the number of used processors [
46
].
Distributed
storing also provides a convenient way to store very large matrices. The
larger
r
U
,
r
I
, and
r
T
are, the more obviously the speedup is.
Note that the user, image, and tag factor matrices are initialized randomly in the
proposed learning algorithm. Likewise to other nonconvex learning problems, the
initialization of the factor matrices is very important to our learning algorithm. We
will be working toward investigating a proper initialization scheme in the future.
2.4 Performance Evaluation
2.4.1 Dataset
We perform the experiments of social tag refinement on the large-scale web image
dataset, NUS-WIDE [
5
]. It contains 269,648 images with 5,018 unique tags col-
lected from Flickr. We crawled the owner information according to the image ID
and obtained the owner user ID of 247,849 images.
8
The collected images belong
to 50,120 unique users, with each user owning about 5 images. We select the users
8
Due to link failures, the owner ID of some images is unavailable.
