Deep learning: Deep neural networks mimic the hierarchical structures in human brains. Their expressive power is exponentially better than shallow models, and has the potential to better capture the complex mapping from low level data features to high level semantics. Deep learning is intrinsically learning a high level representation for the data, and is also called deep representation learning. Deep learning necessitates large-scale training data. However, labeling large-scale training data is very time-consuming. In this direction, I am focused on weakly supervised training of neural networks for multimedia computing using weakly labeled data, to improve the performance of neural networks.

Recommender systems: recommender systems help online users find items (e.g. movies, products) they are interested in. With the rapid development of online social media, the problem of recommendation has been examined in different contexts. I am currently focused on using multi-modal side information (e.g., text, images) for recommendation and developing graph neural networks suitable for capturing complex user behaviors and side information for social media recommendation problems.

Attributed graph mining: real life networks contain not only link structures, but also attributes on vertices and/or edges. For example, in a social network people could buy various products; in a computer network computers could raise different attacking alerts. Studying how different attributes are distributed in a graph is important for revealing the relationships and patterns involving attributes and the graph structure which are hidden in the graph dataset. I have been focused on fundamental correlation mining and ``interesting area\'\' mining in this context, e.g. measuring the correlation between an attribute and the graph structure. I have designed and implemented efficient and scalable tools for assisting people to analyze and mine patterns in attributed graphs. Currently, I am interested in the dynamic aspects of structural correlation.

Medical images + AI: In view of limited medical resources, it is desirable to develop AI models which can diagnose according to patients\' medial imaging data. On the one hand, AI models could learn from top expert doctors and replicate their experiences to lower-level hospitals and clinics; on the other hand, the AI models can also help senior doctors to reduce their workload. In this direction, I am focused on automatic diagnosis of small intestine diseases, facial paralysis and oral mucosal diseases, in collaboration with top hospitals in China such as Ruijin Hospital affiliated to Shanghai Jiaotong University.

E-commerce Advertising: In e-commerce platforms, advertising is a very important component that is beneficial for not only online merchants and the platform, but also the users (improving user experience by showing them proper ads). Thus, it is important to make the advertising system smart. In collaboration with Alibaba group, I am currently investigating two research questions: (1) how to build a bidding system which can help merchants bid user queries automatically and optimize the profits of both merchants and the platform; (2) how to make the results of advertising interpretable.
 

• Key Program of Natural Science Foundation of China, 2020-2024, 3,010,000 rmb, PI
• General Program of Natural Science Foundation of China, 2019-2022, 620,000 rmb, PI
• Program of National Excellent Youth Science Foundation of China, 2016-2018, 1,500,000 rmb, PI
• General Program of Natural Science Foundation of China, 2014-2017, 790,000 rmb, PI