Information Technology Reference
In-Depth Information
2.5 Discussions
In the tag refinement task, we employed the derived factor matrices to analyze the
image
-
tag
associations. As we model the social tagging data by taking into account
all essential entities,
user
,
image
and
tag
, we can apply the model to many other
real-world tasks.
In personalized image search, the returned image results depend on not only their
relevances with the query keywords, but also the relevances with the searchers. For
our case, the associations between users and images can be estimated by measuring
the
user-image
cross-space distances in the same spirits as Eq. (
2.5
), which reflect
the users' preferences and can be leveraged to rerank the returned images.
Another potential application is personalized tag recommendation, whose goal is
to predict tags for each user on a given web item (image, music, URL or publication).
The reconstructed tensor
Y
captures the ternary relationships between users, images,
and tags, where the value of
y
u
1
,
i
1
,
t
1
indicates the likelihood of user
u
1
using tag
t
1
to annotate image
i
1
. Therefore, the tags with the highest
ˆ
ˆ
y
u
,
i
,
t
can be recommended
to user
u
as the potential tags for item
i
.
The proposed RMTF can also be applied to other applications, e.g.,
user profile
construction
and
user recommendation
. It is believed that users express their indi-
vidual interests through tags [
27
], thus the latent user interests can be understood by
estimating the
user
-
tag
association. Actually, we have employed the derived
user
and
tag
factor matrices to build user-specific topic spaces for user modeling, and view the
personalized image search problem as an reexamination into
user
-
tag
-
image
ternary
correlations. Please refer the details to our recent work in [
35
]. Besides exploring the
interrelations, we can directly evaluate the intrarelations among users, images, and
tags in the corresponding subspaces. Users with similar feature representations can
be recommended to each other to connect people with common interests and encour-
age people to contribute and share more content. It is an interesting issue to adapt the
proposed RMTF to more related applications in the future. In addition, there exist
different forms of metadata, such as descriptions, comments, and ratings. While we
focus on tags in this topic, how to model other metadata for a overall understanding
is also our future work.
References
1. Acar, E., Yener, B.: Unsupervised multiway data analysis: a literature survey. IEEE Trans.
Knowl. Data Eng.
21
(1), 6-20 (2009)
2. Bengio, Y., Paiement, J.-F., Vincent, P., Delalleau, O., Roux, N.L., Ouimet, M.: Out-of-sample
extensions for lle, isomap, mds, eigenmaps, and spectral clustering. In: NIPS (2003)
3. Borghol, Y., Ardon, S., Carlsson, N., Eager, D., Mahanti, A.: The untold story of the clones:
content-agnostic factors that impact youtube video popularity. In: Proceedings of the 18th
ACMSIGKDD International Conference on Knowledge Discovery and DataMining, KDD'12,
pp. 1186-1194 (2012)
