Information Technology Reference
In-Depth Information
The first function is to return the users with the highest scores when specifying target
attribute
t
. The second function is to return the users with the highest scores when
specifying auxiliary attributes
s
. The third function is to return the users with the
highest scores when specifying attributes
t
and
ˆ
s
. Actually, the structural attribute-
based user retrieval is analogous to the Facebook's new darling-Graph Search,
10
which is expected to lead the upcoming trend in social networking discovery.
ˆ
3.6 Performance Evaluation
3.6.1 Experimental Setting
We evaluate the effectiveness of the proposed Relational LSVM in three experiments:
attribute relation compatibility analysis, attribute inference, and attribute-based user
retrieval. In all experiments, we use 50% of the labeled data for training and the rest
for testing.
We examine the following methods for comparison in user attribute inference and
attribute-based user retrieval:
•
SVM-Simple: The method using six types of user features respectively to train an
attribute classifier for one attribute type.
•
Stacked SVM: The method applying stacked SVM to the concatenated confidence
scores produced by the attribute classifiers learned with user features.
•
Relational LSVM: The proposed method, which incorporates the attribute relation
for inference.
Mean Average Precision (MAP) is utilized to evaluate the attribute inference
performance. For evaluating the performance of attribute-based user retrieval, we
use Normalized Discounted Cumulative Gain (NDCG), where NDCG at position
k
is defined as:
k
2
r
i
1
IDCG
×
−
1
NDCG@
k
=
(3.15)
log
2
(
i
+
1
)
i
=
1
where
r
i
is the relevance rating of item at rank
i
. For retrieval,
r
i
is 1 if the user has a
link to the returned item and 0 otherwise. IDCG is a normalization constant so that
the perfect ranking has a NDCG value of 1.
The relational SVM model is implemented in Matlab. All experiments are con-
ducted on a PC runningWindows 7with four Intel Core i5-3470 processors (3.2GHz)
and 4GB memory.
