Information Technology Reference
In-Depth Information
Fig. 3.5
The predicted user attributes of the testing sample users for user profiling
3.6.2.3 Attribute-Based User Retrieval
The goal of attribute-based user retrieval is to return the users that match a query
constructed by attributes. To evaluate how our model performs in this task, we create
three sets of queries that consist of one attribute (denoted as Query1), two attributes
(denoted as Query2), and three attributes (denoted as Query3) respectively. The total
number of queries is 562, including 30 queries in Query1, 180 queries in Query2,
and 352 queries in Query3. For Query1, we apply the model corresponding to the
query attribute to score test users for ranking. For Query2 and Query3, each model
corresponding to one attribute from the query is first used to score the test users. Then,
we sum each attribute inference score to indicate the relevance of the candidate users.
For example, given a query of
{“elder, IT person, positive”}, we first apply each
classifier of the attribute in the query on the test set. Then, the response scores are
summed for user retrieval. Sepecially, for our Relational LSVMmodel, we iteratively
select each attribute in the query as target attribute and other attributes as the auxiliary
attributes. Equation (
3.14
) is then used to score the test users. We finally combine the
response scores of each selected target attribute for ranking.
Figure
3.6
depicts the comparison results. We see that our proposed model consis-
tently and significantly outperforms the baselines. It is consistent with the previous
experimental results of user attribute inference. In addition, we can observe that our
model outperforms in more complex cases, i.e., Query2 in Fig.
3.6
b and Query3 in
Fig.
3.6
c. It indicates the significance of attribute relations in attribute-based user
retrieval task, especially when the query consists of multiple attributes.
Figure
3.7
shows three examples of user retrieval results of three queries. The
row represents the retrieval result for one query and the users are displayed in a
descending order by scores. The attributes that match the queries are highlighted in
red, and the incorrect samples are framed in green rectangles. It is shown that most
retrieval results satisfy the queries. For example, in the first row, all 5 returned users
perfect match the query “photographer”. For more complex queries that contains
multiple query attributes, such as the second and third query, most results are correct
Q
=
