Information Technology Reference
In-Depth Information
Fig. 3.3
Illustration of the
proposed model. Each
circle
corresponds to one variable,
and each
square
represents
one factor in the model
the attribute relations to boost the attribute inference capability. A graphical illus-
tration of the proposed model is shown in Fig.
3.3
. At the lowest level, rich user
features are extracted from the multimedia interaction of an individual user. At the
intermediate level, the compatibility between a user's feature vector and attribute
label is modeled. At the top level, attribute relations are explored for accurate target
attribute inference.
3.4.3.1 Relational Latent SVM Model
We now describe how we model the attributes for an individual user. A user
u
is
denoted as a tuple
(
x
,
s
,
t
)
.Here
x
is the user feature vector,
s
=
(
s
1
,...,
s
L
)
∈
S
is
the auxiliary attributes, and
t
is the target attribute. Our goal is to learn a model
that can be used to infer the accurate target attribute label to a test user sample
x
.
To explore the dependent relations between user attributes, we use an undirected
graph
∈
T
G
=
(V , E )
to model user attributes and capture the attribute relations.
Avertex
v
i
∈
V
corresponds to one attribute with certain value, and an edge
(
indicates the relation strength between attribute
a
i
and attribute
a
k
.
We build the graph
i
,
k
)
∈
E
by examining the co-occurrence statistics of user attribute
in the training dataset. The edge weight of attribute
a
i
and attribute
a
k
is defined
as
G
Q
(
a
i
,
a
k
)
ˀ(
,
)
=
(
a
i
,
a
k
)
i
k
, where
Q
denotes the number of co-occurrence
min
(
Q
(
a
i
),
Q
(
a
k
))
(
a
i
)
between attribute
a
i
and
a
k
, and
Q
denotes the number of occurrence for attribute
a
i
. A large weight
w
means a strong relation between attribute
a
i
and
a
k
.
Formally, we are interested in learning a score function
f
w
:
X
×
L
ₒ R
(
i
,
k
)
over an
example user
x
and the target attribute
t
, where
w
are the model parameters. Here
X
denotes the user feature space. We use
f
w
(
to measure the compatibility among
the user feature
x
, the target attribute label
t
, and the configurations of auxiliary
attribute labels
h
.
f
w
(
x
,
t
)
max
s
w
T
x
,
t
)
takes the form of
f
w
(
x
,
t
)
=
ʦ(
x
,
s
,
t
)
to score
