Broad Learning on Heterogeneous Data for Knowledge Fusion

In the era of big data, data are more abundant in their formats and sources with a wide variety of natures and properties. Meanwhile, knowledge discovery tasks are no longer restricted to a single source or homogeneous data. Effective fusion models on heterogeneous data provide opportunities to understand the information entities more comprehensively. However, the variety of big data brings challenges, such as sparsity of labels, the difficulty of network alignment and synchronization problem, to model design. Broad learning is a learning framework aiming at knowledge fusion on heterogeneous data addressing the above challenges. In this dissertation, I will introduce my research on principles, models and algorithms of broad learning in several different aspects. First, we proposed a transfer learning model to solve the label sparsity problem when fusing knowledge from multiple domains. Second, we proposed the online consensus maximization model to address the challenge of synchronization when fusing the power of multiple predictive models. Third, I will introduce our temporal fusion model, learning from multiple subgraphs in data streams. Finally, I will present our fusion model on heterogeneous information networks, a deep framework for missing entity completion and ranking in online knowledge libraries.