Local Patch Framework for Fixing Supervised Learning Models
I worked from June 2011 to March 2012 as a research intern in Machine Learning Group, Microsoft Research Asia (MSRA), under the guidance of my mentor, Jun Yan, a lead researcher in MSRA. We propose a novel local patch framework to break the black box of supervised learning models by identifying and fixing the “local bugs” existed in learning models and generalize the patched model on future unseen data. In this local patch framework, Metric Learning technique is utilized to minimize the impact on original success cases and focus on fixing the local region in data space which may contain most of the failure cases. The proposed local patch framework shows good performance on experiments in both classification and ranking problems and largely outperforms other globally model fixing solutions.
Multi-Objective Optimization (MOO) Framework in Sponsored Search
The second research topic I have done in MSRA is Multi-Objective Optimization framework in sponsored search. The goal of this research is to balance the utilities of users, advertisers and publishers in sponsored search to help search engine improve long term revenue and provide effective prediction of tradeoff parameters for those participants using machine learning algorithms.
Quality Control of SVC Video Streaming
From June 2010, I have been doing research in Peking University Undergraduate Research Program, in Institute of Computer Science and Technology of Peking University, under the guidance of Jun Sun, an associate professor in PKU. I have two tasks in this program.
One is researching on quality control for Scalable Video Coding (SVC) which aims to reduce video quality fluctuation. I propose a quality control algorithm based on Proportional-Integral–Derivative (PID) control concept. This algorithm enables continuous monitoring and predicting of real-time bandwidth. It provides a more flexible mechanism to adjust the video quality so that it has a more accurate and swift response to the changing network environment, delivering online SVC video streaming at the best quality. This quality control algorithm has reduced quality fluctuation significantly to improve audience satisfaction.
The other task is to develop a P2P system to transport SVC which will be discussed in detail in "Projects" section.
Character String Matching Optimization Algorithms Based on BM
When I was a student in "Data Structure and Algorithms(A)(Honor Track)" in autumn, 2009, I found the deficiency in BM string matching algorithm. Then I did a short research under the guidance of Ming Zhang, a professor in PKU EECS. This research tried to reduce the complexity of BM algorithm in its worst situation. The advantage of my proposed algorithm has been proved by experiment and can be applied in Gene Match engineering.
Maximum Clique Problem
I have once done a group research on max clique problem as a course project under the guidance of Kaile Su, a professor in PKU EECS. Max clique problem has long been treated as the classical problem to illustrate NP hard problem in Computer Theory. Based on analysis of current algorithms, we proposed new local search algorithm and tested influence of some parameters. Our algorithm reached an achievement that was top three in world in frb100-40 challenge. (Technical Report)