Wang, Yingze
(2014)
LEARNING WITH SPARSITY FOR DETECTING INFLUENTIAL NODES IN IMPLICIT INFORMATION DIFFUSION NETWORKS.
Doctoral Dissertation, University of Pittsburgh.
(Unpublished)
Abstract
The diffusion of information and spreading influence are ubiquitous in social networks. How to model and extract useful information from diffusion networks especially in social media domain is still an open research area that requires significant attention. Many real applications pose new challenges in modeling information diffusion process. In particular, the first challenge comes from the fact that the underlying network structure over which the propagation spreads is unknown or unobserved. It is often the case that one can only observe that when nodes got infected by which contagion but without the knowledge about who infecting whom. The second challenge comes from the simultaneous transmissions of multiple correlated contagions through an implicit network. The third one comes from strong temporal effect in the diffusion process which needs to be carefully modeled.
In my thesis, we address two fundamental tasks, forecasting and influential-node detection, in an implicit diffusion network by a unified approach. In particular, we first proposed a sparse linear influence model (SLIM) which takes a nice form of a convex optimization problem. We further extended SLIM to multi-task sparse linear influence model (MSLIM), which could model diffusion networks with multiple correlated contagions. MSLIM, as a richer model than SLIM, not only improves prediction accuracy, but also allows to select influential nodes on a finer grid, i.e., select different sets of influential nodes for different contagions. For SLIM and MSLIM, we developed both deterministic and stochastic optimization algorithms for solving the corresponding problems and showed the fast theoretical convergence guarantees.
Another contribution of the thesis is the development of a general purpose system, called Slow Intelligent System (SIS), which is able to continuously learn and improve performance over time. We proposed the component-based SIS and developed the software with applications to face recognition task. Furthermore, we utilized the idea of the SIS to systematize the information diffusion process modeling and influential node detection and proposed SIS-based SLIM/MSLIM approaches, which further improve the flexibility and scalability of learning from implicit diffusion networks. We demonstrated the superiority of the proposed approaches on several real datasets from social media domains.
Share
Citation/Export: |
|
Social Networking: |
|
Details
Item Type: |
University of Pittsburgh ETD
|
Status: |
Unpublished |
Creators/Authors: |
|
ETD Committee: |
|
Date: |
30 May 2014 |
Date Type: |
Publication |
Defense Date: |
26 March 2014 |
Approval Date: |
30 May 2014 |
Submission Date: |
15 April 2014 |
Access Restriction: |
No restriction; Release the ETD for access worldwide immediately. |
Number of Pages: |
135 |
Institution: |
University of Pittsburgh |
Schools and Programs: |
Dietrich School of Arts and Sciences > Computer Science |
Degree: |
PhD - Doctor of Philosophy |
Thesis Type: |
Doctoral Dissertation |
Refereed: |
Yes |
Uncontrolled Keywords: |
Social Network Analysis, Slow Intelligence System, Information Diffusion Network, Machine Learning, Sparse Learning, Optimization, Regression, Multi-task Regression, Text Mining |
Date Deposited: |
30 May 2014 15:03 |
Last Modified: |
15 Nov 2016 14:19 |
URI: |
http://d-scholarship.pitt.edu/id/eprint/21246 |
Metrics
Monthly Views for the past 3 years
Plum Analytics
Actions (login required)
|
View Item |