My name is Xun Zhao, a Ph.D. student in the Department of Computer Science and Engineering at Hong Kong University of Science and Technology (HKUST). I am supervised by Prof. Huamin Qu and Prof. Dik Lee . I obtained my B.Sc. degree in HUST (Hua Zhong University of Science and Technology) majored in Automation.
My research interest lies in the intersection of Machine Learning, Data Visualization, and Human-Computer Interaction.
My research interests lie in the intersection of Machine Learning, Data Visualization and Human-Computer Intersection.
Skyline query is widly used to automatically selecting superior data points for decision making. However, these skyline points are difficult for people to compare and interpret. We present a visual analytic system aiming at revealing the superiority of skyline points from different perspectives and at different scales to aid users in their decision making. More details can be found here.
A critique problem for training deep learning network is prevent overfitting while keep the model capacity. We regularize the model by pruning the unimportant connections and neurons. We test the performance on different networks like CNN, RNN and LSTM on the tasks of image classification.
Most recommendation systems rely on user profiles and use fully automated recommendation techniques. They are focused on the perfomance of algorithms and the accuracy is tested offline. However, the goal of recommentation system is introducing users to interesting items and convincing them to try/buy the recommended items. A mismatch exits between the user need and the metric, whichs yarns for engaging the user into the reccommending process.
In order to provide better interest to user, we analyze user twitter history and build user model. We propose a Dynamic Interest Model, which extracts temporal features from twitter history and apply Gaussion Mixture Model to soft-cluster interests into four interest types.We then use interest type to predict the future recurrence of an interest, and give topic recommendations.
We implement the learning of feedforward ANN under Bayesian regularization. We can predict the closing price of next day within error of 2% on average by the MapReduce-based ANN model.
The high-Dimensional property of natural images makes the analysis complicated and diffcult. We decompose natural images of high-dimensional statistical properties to low-dimensional.