Information Technology Reference
In-Depth Information
Table 3.6 Performance comparison of different methods for user attribute inference in terms of
mAP
Examined methods Age
Gender Relationship Occupation Interest Sentiment orientation
SVM-face
0.6194 0.7607 0.5835
0.0741
0.5005 0.3398
SVM-profilephoto
0.5422 0.7185 0.5181
0.0776
0.5002 0.3579
SVM-postphoto
0.5047 0.6276 0.5193
0.1098
0.5215 0.3671
SVM-unigram
0.5989 0.7239 0.5899
0.2329
0.5490 0.4002
SVM-sociolinguistic 0.5972 0.7123 0.6081
0.2002
0.5501 0.3922
SVM-topic-based
0.5264 0.5768 0.5376
0.0798
0.5037 0.3333
Stacked SVM
0.6054 0.7856 0.6114
0.2373
0.5980 0.4096
Relational LSVM 0.7278 0.7986 0.6240
0.2507
0.6172 0.4106
interest. That means when training a model for the value “technology”, for instance,
we use all user samples containing the “technology” label to construct the positive
set and the rest for the negative set.
Table 3.6 lists the inference results of different methods, from which we observe
that (1) The types of user features contribute differently to user attribute infer-
ence. For example, the profile photo and face features are rather powerful when
inferring attribute “age” and “gender”; The term presence based SVM-unigram
and SVM-sociolinguistic model perform better than other feature-based models in
terms of “interest”, “occupation” and “sentiment orientation” attribute inference. This
verifies the effectiveness of our discriminative term selection strategy. Concept-based
image representation for post photos and LDA topic-based feature for text posts show
poor performance. It illustrates that it is challenging to extract a compact and dis-
criminative feature representation at user-level from post content. (2) Stacked SVM
consistently outperforms SVMs based separately on the 6 user features in all tested
cases. It confirms the superiority of stacked SVM in combing different types of user
features for attribute inference. (3) Our proposed Relational LSVM model incorpo-
rating user attribute relations significantly boosts the performance of user attribute
inference comparing with other baselines. Particularly, Relational LSVM improves
the age attribute inference over stacked SVM by 12.24%. The performance of other
attribute inference is also improved. This result validates the advantage of exploiting
attribute relations for user attribute inference.
In Fig. 3.5 , we present some attribute inference to construct the user profiles by
the Relational LSVMmodel. The model accurately infers most of the user attributes,
such as target attributes (“male”, “young”, and “unmarried”) and even auxiliary
attributes (“elder” and “IT person”). However, as discussed previously, we also note
that auxiliary attribute inference shows more incorrect instances than that of target
attribute inference.
 
Search WWH ::




Custom Search