In machine learning area, we have developed a suite of theory and techniques on multi-task learning and transfer learning. Specifically, we have developed novel theory and algorithms leading to effective solutions to the challenging scenarios of multi-task learning where there are tasks with very few training samples and transfer learning where there is no training data at all in the target domain. We have applied the theory and techniques to solve for the challenging application in imperfect imagery/video annotation where the training samples may be error-prone and incomplete, which is a very realistic scenario in the real-world. We are the first and still in the leading-edge in the literature to address this challenging problem.
In data analytics area, we have developed a suite of theory on multimedia data concept discovery and a suite of theory on relatoinal data clustering. We have published the very first monographs on these two topics, respectively. As a special case of relational data, we have developed a novel theory as well as the related algorithms for citation networks, which are special relational data with a wide spectrum of applications in the real-world such as Web data, emails, and scientific publications. Our theory and techniques have found many important applications including social network analysis, bioinformatics, e-commerce, geointelligence, as well as intelligent recommenders. We have recently also developed an effective link prediction method for sparse graphs using a combination of matrix factorization and autoencoder through dropout training.
As another specific application of our theory on relational data learning related to Internet of Things, we have developed effective graph partitioning methods that can be used to a network of sensors for prediction of events such as fluding or earthquake.
The research in these areas is published in my lab in the top-notch venues including ICML, ACM KDD, IEEE ICDM, SDM, AAAI, ACM MM, IEEE ICME, and ICPR.
Please visit my webpage for the relevant information and the links to the publications.