THESIS
2014
xiv, 162 pages : illustrations ; 30 cm
Abstract
People now live in a social world. The advent of online social networks not only
connect us more tightly, but also enables us to record and share our daily lives at a fast
rate. The mobile devices such as smartphones and smartglasses help people to connect
the online social world and the physical world. Activity recognition, which aims to
infer the intentions and goals of the user, can enhance user experience of online social
networks and wearable devices by providing better and more intelligent services. In
this thesis, we study a special kind of activity recognition problems in which external
social knowledge is available. The social knowledge considered in this thesis consists
two kinds: social network structure properties and abundant activity records in online
social netwo...[
Read more ]
People now live in a social world. The advent of online social networks not only
connect us more tightly, but also enables us to record and share our daily lives at a fast
rate. The mobile devices such as smartphones and smartglasses help people to connect
the online social world and the physical world. Activity recognition, which aims to
infer the intentions and goals of the user, can enhance user experience of online social
networks and wearable devices by providing better and more intelligent services. In
this thesis, we study a special kind of activity recognition problems in which external
social knowledge is available. The social knowledge considered in this thesis consists
two kinds: social network structure properties and abundant activity records in online
social networks. Utilizing these two knowledge sources has unique challenges. The
first challenge is to utilize the the social structure property and incorporate the influence
among social users. The second challenge is the data sparsity problem in which the
labeled data is limited due to reasons such as short usage time, inactiveness of some
users and high cost. The third challenge is the difference in feature spaces when we
have activity records from both physical world and online virtual world.
We develop a unified transfer learning framework which can effectively solve the
above three challenges and therefore improve activity recognition accuracy. The unified
framework can support two different styles of knowledge transfer: feature learning and model learning. Under this general learning framework, we generalize existing transfer
learning methods and develop new ones. In particular, we develop a Heterogenous
Transfer Learning model that can transfer social knowledge across feature spaces. To
illustrate the effectiveness and generality of the framework, we apply four specific
models derived from it to four representative activity recognition applications: social
spammer detection, social activity level prediction, semantic place prediction, and
heterogeneous transfer from online social activities to the physical world. Our experimental
results on the four specific recognition tasks all demonstrate the high effectiveness of
the proposed transfer learning framework.
Post a Comment