In recent years, micro-video apps have rapidly emerged and become important channels for entertainment and socializing.
For micro-video platforms, the number of active users is the gold standard for measuring the success of the platform. In addition to attracting, motivating, and cultivating outstanding creators, the optimization of recommendation algorithms based on artificial intelligence (AI) technology is also a very important part.
Boosting micro-video recommendation performance via exploring and modeling the explicit correlations among multi-modalities is one of the research directions of Dr. Liqiang Nie with Shandong University. His research interests primarily lie in multimedia analysis and search, aiming to resolve two key research problems: explicit fusion of correlated multimodalities in feature learning and guiding models towards semantic understanding. He has proposed a series of data-driven multimodal fusion and representation learning models that fill the theoretical gap in data-driven multimodal fusion.
Liqiang graduated from Xi'an Jiaotong University and went to the National University of Singapore (NUS) to pursue his Ph.D. under the supervision of Professor Chua Tat-Seng, the founding Dean of the School of Computing at NUS. Over the past eight years, Liqiang first obtained his Ph.D.in 2013, and then worked as a research assistant and research at NUS sequentially, focusing on data-driven multimodal learning.
At the end of 2016, Liqiang returned to China and joined Shandong University as a full professor, and established the Intelligent Media Research Center (iLearn) of Shandong University to expand his research scope.
Liqiang believes that knowledge cannot only improve the performance of the model and reduce the dependence on labeled data but also teach or guide the model to think and to reason about things that have not been seen before. So, he led the team to put forward the concept of guiding multi-modal reasoning based on prior knowledge and rules.
He designed and verified a series of original multimodal reasoning approaches such as hierarchy-regularized multimodal hashing, graph-constrained multimodal multitask learning and multimodal learning with probabilistic knowledge distillation. These models greatly alleviate the data dependence and strengthen the model generalization and reasoning abilities.
Besides video recommendation, his work on data-driven multimodal learning and knowledge-guided multimodal reasoning can also be extended to application scenarios such as power grid, clothing matching, and dialogue robots, providing significant economic and social benefits.
In the future, Liqiang plans to engage in demand-driven multi-view sensing and multimodal edge computing, and further promote the development of intelligent operation and maintenance of power grid.