Information Technology Reference
In-Depth Information
3.5.1 User Profiling
User profiling is the process of obtaining values of different properties that constitute
the user model. Methods for user profiling typically fall into two categories: profile
extraction and profile learning. Our model can be directly applied to user profiling.
Given a new user feature
x
, we ca use the model to infer the target attribute and
auxiliary attribute as:
max
s
t
∗
=
w
T
ʦ(
argmax
t
x
,
s
,
t
)
(3.13)
s
∗
=
w
T
t
∗
)
argmax
s
ʦ(
x
,
s
,
We note that selection of the target attribute and the auxiliary attribute depends on
the application scenario. For example, if interested in predicting user's occupation,
we treat the occupation attribute as the target attribute and rest attributes as auxiliary
ones. The difference between the prediction of target attribute and auxiliary attribute
is that the model tends to guarantee more accurate predictions for the target attribute
comparing with auxiliary attributes. The reason is that the optimization objective
function in Eq. (
3.7
) aims to maximize the margin for target attribute classification.
To remedy this limitation, we conduct user profiling by setting each attribute as the
target one and run the Relational LSVM model iteratively. After attribute inference,
we can construct the user profile and describe the user with multiple attributes.
3.5.2 Attribute-Based User Retrieval
Our model can also be used for tackling the retrieval task. Recall that we employ
the graph
to encode the dependency relations for the attribute prediction model.
After learning the model, the parameter
G
w
encodes certain relation com-
patibility between attributes. The user attribute relation compatibility can be utilized
to assist user retrieval. For example, the male users that are interested in information
technology are more probably to be returned when searching for IT related users.
Based on the model, we extend the attribute-based user retrieval to enable the queries
constructed by one or more attributes and return the most relevant users correspond-
ing to the specified attributes. Formally, we cast the task into solving the following
optimization problem:
{
ʳ,ʷ
}ↂ
x
∗
=
w
T
ʦ(
argmax
s
x
,
s
,
t
),
given
t
.
max
s
x
∗
=
w
T
s
,
ˆ
argmax
t
ʦ(
x
,
s
,
t
)
,
given
s
ˆ
.
(3.14)
x
∗
=
w
T
s
,
ˆ
ʦ(
argmax
s
x
,
s
,
t
),
given
t
,
ˆ
s
.
