Information Technology Reference
In-Depth Information
the confidence of user sample
x
labeled as target attribute label
t
with the auxiliary
attribute configuration
s
, which is defined by combining different potential functions:
w
T
T
T
ʦ(
x
,
s
,
t
)
=
ʱ
ˆ(
x
,
t
)
+
ʲ
˕(
x
,
s
i
)
(3.2)
i
∈
S
T
T
+
ʳ
ˉ(
s
i
,
t
)
+
ʷ
ˈ(
s
i
,
s
k
,
t
)
∈
S
)
∈
E
i
(
i
,
k
See Fig.
3.3
for an illustration of our model. In this model the parameter vector
w
is the concatenation of the parameters in all factors. The model simultaneously
considers the dependencies among user features and attributes. In particular, the first
term predicts the target attribute label from user features; the second term describes
the relationship between user features and auxiliary attributes; the third term captures
the relationship between auxiliary attributes and target attribute; and the last term
represents the dependencies between auxiliary attributes. The details of potential
functions in Eq. (
3.2
) are described in the following.
User Feature Versus Target Attribute Potential
T
ʱ
ˆ(
x
,
t
)
: This potential is a stan-
dard linear model for target attribute prediction. Here
represents a certain
mapping of the user feature
x
and the mapping result depends on attribute label
t
.
In the implementation, we represent
ˆ(
x
,
t
)
ˆ(
x
,
t
)
as the score of the target attribute SVM
classifier. Model parameter
re-weights the SVM scores to make the target attribute
prediction more accurate. Specifically, the potential is parameterized as:
ʱ
T
ʱ
ˆ(
x
,
t
)
=
ʱ
b
1
·
(
t
=
b
)
·
g
b
(3.3)
b
∈
T
where
is the indicator function,
g
b
is the SVM confidence score corresponding
to target attribute
b
.
User Feature Versus Auxiliary Attribute Potential
1
(
·
)
T
: This potential is a
standard linear model trained to predict the value of the
i
th auxiliary attribute given
the user feature
x
. Similar to the potential in user feature versus target attribute model,
˕(
ʲ
˕(
x
,
s
i
)
represents the confidence score from the auxiliary attribute SVM classifier.
The parameter
x
,
s
i
)
re-weights the auxiliary attribute SVM scores. This potential func-
tion is parameterized as:
ʲ
T
ʲ
˕(
,
s
i
)
=
ʲ
c
·
1
(
=
)
·
x
s
i
c
y
c
(3.4)
c
∈
S
where
y
c
is the SVM score corresponding to auxiliary attribute
c
.
Target Attribute Versus Auxiliary Attribute Potential
T
: This potential
represents the dependencies between the target attribute
t
and auxiliary attribute
s
i
. Model parameter
ʳ
ˉ(
s
i
,
t
)
encodes the attribute relation compatibility between target
attribute and auxiliary attribute. The potential function is defined as:
ʳ
