SCALABLE ALGORITHMS FOR HIGH DIMENSIONAL STRUCTURED DATA

Abstract

Emerging technologies and digital devices provide us with increasingly large volume of data with respect to both the sample size and the number of features. To explore the benefits of massive data sets, scalable statistical models and machine learning algorithms are more and more important in different research disciplines. For robust and accurate prediction, prior knowledge regarding dependency structures within data needs to be formulated appropriately in these models. On the other hand, scalability and computation complexity of existing algorithms may not meet the needs to analyze massive high-dimensional data. This dissertation presents several novel methods to scale up sparse learning models to analyze massive data sets. We first present our novel safe active incremental feature (SAIF) selection algorithm for LASSO (least absolute shrinkage and selection operator), with the time complexity analysis to show the advantages over state of the art existing methods. As SAIF is targeting general convex loss functions, it potentially can be extended to many learning models and big-data applications, and we show how support vector machines (SVM) can be scaled up based on the idea of SAIF. Secondly, we propose screening methods to generalized LASSO (GL), which specifically considers the dependency structure among features. We also propose a scalable feature selection method for non-parametric, non-linear models based on sparse structures and kernel methods. Theoretical analysis and experimental results in this dissertation show that model complexity can be significantly reduced with the sparsity and structure assumptions.