Information Technology Reference
In-Depth Information
Table 3.5
Detailed concept list
Concept list
animals beach beauty bird bodypart topics building car cartoon cat celebrity child city cityscene
cloth cloud colors couple crowd dancing dark design dog drink electronic product family flight
flower food fruit geek goose grass grassland house icon indoor insect lake landscape leaf man
model mountain naturescene office painting palace party people performance photography portrait
poster puzzle road room sculpture sea sky snow soldier sport spot squirrel stadium stone store street
sunset talk text tiny plant tower toy transport tree universe watch waterfall woman
is manually constructed based on the observation of 88,988 downloaded post photos.
The reason we construct the concept list via manually defining instead of automati-
cally learning is two-fold: (1) By investigating into the post photos, we observe some
common concepts, e.g., people, animals, birds, snow, lake, and mountain. We can
aggregate the photos that describe a common concept for training; (2) The concepts
learned automatically capture less semantics and are difficult to interpret. We select
81 categories from the post photos to construct the concept list, which is shown in
Table
3.5
. Each concept category contains around 100 photos for training. We train 81
concept classifiers in a supervised manner. Dense HOG features [
9
] are extracted for
each image and Locality-constrained Linear Coding (LLC) [
31
] is utilized to obtain
the image representation. For each concept, we train a SVM classifier using LIB-
LINEAR [
11
]. The classification confidence is mapped to a probability score by the
sigmoid function. Therefore, each photo is finally represented as an 81-dimensional
vector corresponding to the concept probability score. Since a user may post more
than one photo, we apply a max-pooling method to aggregate multiple photo feature
vectors and finally obtain an 81-dimensional feature vector for each user, which is
referred as the post photo feature.
3.4.2 Stack SVM-Based User Attribute Inference
For user attribute inference, we learn predictive models based on the extracted user
features. Particularly, for each user attribute, we build six SVM classifiers based on
the six types of user features using LIBLINEAR [
11
], respectively.
To derive the attribute value from six classifiers, a fusion scheme is desired to
combine the six confidence scores. We employ a stacked model [
32
] to perform the
fusion. The key idea of stacked model is to learn a meta-level (level-1) classifier
based on the output of base-level (level-0) classifiers. Although simple, this scheme
well solved our problem of integrating multiple feature-base classifiers.
3.4.3 Exploring Attribute Relation for User Attribute Inference
It is intuitively recognized that user attribute correlates with each other for the same
individual. Therefore, we develop a structural discriminative model for incorporating
