Broad Learning in Multiple Heterogeneous Domains

Recent years have witnessed the flourishing of heterogeneous data from various types of domains. For example, online review sites (like Amazon and Yelp) have access to contextual information of shopping histories of users, the reviews written by the users, as well as the description of the items. Broad learning is introduced to fuse such rich and complex heterogeneous information to improve the performance of learning tasks at hands. In this dissertation, I will introduce our latest research progress on broad learning in multiple heterogeneous domains. In the first part, I focus on heterogeneous network based approaches for connecting and transferring knowledge across domains. To analyze the hidden connections and correlations between different domains, I present a methodology for identifying the same users in multiple domains. I further propose a personalized recommendation algorithm that utilizes complementary information from related domains to improve recommendation performance. To model the complex multi-way relationships among multiple tasks, in the second part, I propose a tensor-based framework for learning the predictive multilinear structure to solve multiple tasks altogether. Moreover, I present a generic method for learning structural data from heterogeneous domains, which can efficiently explore the high order correlations underlying relational structures of multi-way interactions.