Information Technology Reference
In-Depth Information
Fig. 3.2
The framework
six types of user attributes:
gender, age, relationship, occupation, interest,
and
sen-
timent orientation
. Each type of attribute has multiple values. Two challenges are
involved during the attribute inference: (1) how to exploit the heterogeneous and
multimodal user interaction for user attribute derivation? and (2) how to explore the
dependency relations between different types of user attributes for more accurate
attribute inference?
We propose a Relational Latent SVM (Relational LSVM) model-based frame-
work to address the challenges. In particular, we take Google
+
, the popular social
network sites, as the test platform. In Google
, users are allowed to build their pro-
files on the
About
board and post activities on the
Posts
page. As shown in Fig.
3.2
,we
formulate the relational user attribute inference problem as follows: the input is user's
social networking information including profiles from
About
and posts activities from
Posts
in Google
+
. The Relational LSVM model is developed to learn the output in
a supervised discriminative fashion, including the predicted user attributes and the
inferred attribute relations. Within the model, one type of user attribute is selected
as
target attribute
and the remaining are treated as
auxiliary attributes
. While
Ta rge t
attribute
obtains direct reinforcement, the auxiliary attributes are treated as latent
variables during the inference. Multiple relations between auxiliary attributes and
target attribute are jointly formulated as potential functions into the model and help
refine the inference of target attribute. With the derived user attribute and attribute
relations, we apply them to applications of user modeling and attribute-based user
retrieval. We evaluate the model on a collected real-world dataset with full attribute
annotations from Google
+
. The results demonstrate the effectiveness of our Rela-
tional LSVMmodel for user attribute inference and the potential of attribute relation
in user-centric applications.
+
