SPARSE STATISTICAL LEARNING FOR HIGH-DIMENSIONAL TEXT MINING AND TIME SERIES FORECASTING

Abstract

Recent advances in sparse learning techniques have attracted attention by providing new tools to understand, explore and evaluate high dimensional data in many areas. Under the assumption of sparsity, a parsimonious and compact model is preferred, where the sparse learning methods could lead to more consistent models with both better theoretical properties and empirical results. In this thesis, we develop structured sparse learning methods called Regularized Text Logistic (RTL) regression model along with scalable optimization algorithms to model high-dimensional data with complex structures and dynamics. We also propose a sparse geometric regularization framework called SWaGEL, a composite scheme that combines Sparse Wasserstein distance with Group and Exclusive Lasso together, applied in text mining and computer vision. Lastly, we introduce a transformational journey on applying sparse time series forecasting and inventory optimization techniques to limousine scheduling problem in a major hotel in Singapore. We propose a statistical learning approach to optimize the schedule of the hotel and to support decision-making for planners and controllers to drive sustained business value. This has resulted in up to S$3.2 million of savings per year while improving the service level.