huankaip_ECE_2016.pdf (4.24 MB)
Understanding and Engineering Social Dynamics
thesis
posted on 2016-02-03, 00:00 authored by Huan-Kai PengAll activities on social media evolve with time. Consequently, being able to understand
and engineer social dynamics, the way how various properties of social media
evolve, is a central question for social networks research. While recent work has studied
social dynamics from various angles, two important properties of social dynamics
are yet to be addressed, i.e., heterogeneous features and signatures at multiple time
scales. However, considering heterogeneous features is necessary to build a general
tool with wide applicability, whereas considering multiple time scales is indispensable
to study how social dynamics in dierent time scales interact with one another.
In this thesis, we aim at addressing these two properties using computational
algorithms with statistical groundings. In particular, we propose scalable and eective
methods for three basic tasks: pattern mining, structure decomposition, and datadriven
dynamics engineering. For each task, the proposed methods are analyzed
formally and veried empirically. The results reveal several interesting insights and
demonstrate various practical applications, such as dynamics prediction, anomaly
detection, and targeted intervention. Finally, the methods we propose in this thesis
are general enough to handle multi-dimensional time series; we have explored this
direction by considering other applications, such as human behavior recognition and
macroeconomics.
and engineer social dynamics, the way how various properties of social media
evolve, is a central question for social networks research. While recent work has studied
social dynamics from various angles, two important properties of social dynamics
are yet to be addressed, i.e., heterogeneous features and signatures at multiple time
scales. However, considering heterogeneous features is necessary to build a general
tool with wide applicability, whereas considering multiple time scales is indispensable
to study how social dynamics in dierent time scales interact with one another.
In this thesis, we aim at addressing these two properties using computational
algorithms with statistical groundings. In particular, we propose scalable and eective
methods for three basic tasks: pattern mining, structure decomposition, and datadriven
dynamics engineering. For each task, the proposed methods are analyzed
formally and veried empirically. The results reveal several interesting insights and
demonstrate various practical applications, such as dynamics prediction, anomaly
detection, and targeted intervention. Finally, the methods we propose in this thesis
are general enough to handle multi-dimensional time series; we have explored this
direction by considering other applications, such as human behavior recognition and
macroeconomics.
History
Date
2016-02-03Degree Type
- Dissertation
Department
- Electrical and Computer Engineering
Degree Name
- Doctor of Philosophy (PhD)