Latent Variable Models at Scale: Modeling, Regularization,

One central task in machine learning (ML) is to extract hidden knowledge and structure from observed data. This task is especially valuable with the prosperity of big data, where the data itself is of limited value unless knowledge can be uncovered therefrom. Latent variable models (LVMs) elegantly fit into this task where the observed variables are used to model data and the latent variables are utilized to characterize the hidden knowledge, and have been widely used in text mining, computer vision, computational biology, recommender system, to name a few. In the realm of big data, the complexity of knowledge and the size of data are way beyond those in small scale data analytics, presenting various challenges for latent variable modeling. In my thesis work, I aim to design, regularize and solve LVMs to effectively and efficiently extract knowledge and structure from big data. Research Summary My research targets several key issues in large scale latent variable modeling, including model design, regularization, scalability and privacy, and has proposed LVMs to address various modeling issues in big data analytics, such as integrating correlated tasks [12,8], handling multi-modal data [13], incorporating external knowledge [7,6]; A diversity regularization approach to effectively capture long-tail factors in latent knowledge [1] an