The Hong Kong University of Science and Technology
Department of Computer Science and Engineering
PhD Thesis Defence
Title: "Activity Recognition via Social Knowledge Transfer"
Mr. Yin ZHU
In today's world, we have increasingly sophisticated means of recording the
daily activity of humans as well as other moving objects in both physical and
virtual worlds. These recorded activities include phone calls, uses of Apps on
smartphones, and expression of opinions over social-network objects such as
photos. These activities and actions give rise to a huge amount of data.
Activity recognition aims to understand users' actions and intentions based on
models built from these data. Accurate activity recognition allows us to track
people's daily activities, to recognize the semantic functions of places, to
help users and managers identify spammers in an online social network, and to
predict the future activity levels of social-network users.
In this thesis, we study a special kind of activity recognition problems in
which a social network structure is available or abundant social activity
records are given as external knowledge sources. We call these two kinds of
auxiliary knowledge sources as social knowledge. Utilizing them has two unique
challenges. First, how to utilize the rich knowledge in the social structure
and model it for the activity recognition model for each user in the social
network. Second, while the data is abundant in general, sometimes the labeled
data is limited due to reasons such as short usage time and inactiveness of
some users and high cost to label data in both physical and virtual worlds.
Transfer learning is a new learning framework that is suitable to model
external knowledge sources and becomes especially effective when the target
problem domain suffers from data sparsity issues. In the past decade, transfer
learning has been successfully applied to application domains such as text
mining, image understanding, and recommender systems. We propose a novel
transfer learning framework that uses auxiliary social knowledge to improve
activity recognition tasks, and apply it to four specific problems: social
spammer detection, semantic place prediction, social activity level 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.
Date: Monday, 21 July 2014
Time: 9:30am - 11:30am
Venue: Room 3501
Chairman: Prof. Bertram Shi (ECE)
Committee Members: Prof. Qiang Yang (Supervisor)
Prof. Raymond Wong
Prof. Dit-Yan Yeung
Prof. Rong Zheng (ISOM)
Prof. Irwin King (Comp. Sci., & Engg., CUHK)
**** ALL are Welcome ****