The New Jersey Institute of Technology's
Electronic Theses & Dissertations Project

Title:	Statistical learning methods for mining marketing and biological data
Author:	Zhang, Jie
View Online:	njit-etd2017-056 (xii, 79 pages ~ 1.6 MB pdf)
Department:	Department of Computer Science
Degree:	Doctor of Philosophy
Program:	Computer Science
Document Type:	Dissertation
Advisory Committee:	Wei, Zhi (Committee chair) Calvin, James M. (Committee member) Roshan, Usman W. (Committee member) Wang, Antai (Committee member) Zhao, Zhigen (Committee member)
Date:	2017-05
Keywords:	Change-point detection Gender prediction Sparse time series analysis Multi-task learning Empirical Bayes method Data mining
Availability:	Unrestricted
Abstract:	Nowadays, the value of data has been broadly recognized and emphasized. More and more decisions are made based on data and analysis rather than solely on experience and intuition. With the fast development of networking, data storage, and data collection capacity, data have increased dramatically in industry, science and engineering domains, which brings both great opportunities and challenges. To take advantage of the data flood, new computational methods are in demand to process, analyze and understand these datasets. This dissertation focuses on the development of statistical learning methods for online advertising and bioinformatics to model real world data with temporal or spatial changes. First, a collaborated online change-point detection method is proposed to identify the change-points in sparse time series. It leverages the signals from the auxiliary time series such as engagement metrics to compensate the sparse revenue data and improve detection efficiency and accuracy through "smart" collaboration. Second, a task-specific multi-task learning algorithm is developed to model the ever-changing video viewing behaviors. With the 1-regularized task-specific features and jointly estimated shared features, it allows different models to seek common ground while reserving differences. Third, an empirical Bayes method is proposed to identify 3' and 5' alternative splicing in RNA-seq data. It formulates alternative 3' and 5' splicing site selection as a change-point problem and provides for the first time a systematic framework to pool information across genes and integrate various information when available, in particular the useful junction read information, in order to obtain better performance.