DIR LABORATORY - Key Persons

Dr. Bingchen Yu

Job Titles:

Research Scientist @ Georgia State University

Dr. Lei Li

Job Titles:

Faculty Advisor & Members

Biographical Sketch Meng Han currently is an assistant professor in the College of Computing and Software Engineering at Kennesaw State University. He got his Ph.D. in Computer Science from Georgia State University. He is currently an ACM member, an IEEE member, and an IEEE COMSOC member. His research interests include Data-driven Intelligence, Big Social Data Mining, Cyber Data Security & Privacy, IoT Data System, and Blockchain Technologies, etc.

Dr. Qinlong Luo

Job Titles:

Data Scientist from IBM
Global Chief Data Office, IBM
Industrial Research Scientist

Dr. Qinlong Luo is a Data Scientist from IBM. He got his PhD in Computational Physics from University of Tennessee at Knoxville. He is interested in applying Physics technologies and concepts into Machine Learning and Data Science. The research projects Dr. Luo has been working on includes: Markov Chain Monte Carlo (MCMC) and Clustering: MCMC is widely used in computational physics as an optimization method to solve physics models. Clustering is one kind of unsupervised learning in machine learning, and some of the clustering problems can be solved by optimization. As an optimization method, MCMC can be applied to a variety of clustering problems. For instance, Word Sense Disambiguation (WSD) is one of the most important topics in the field of Natural Language Processing. In order to figure out how many different meanings an ambiguous word can refer to automatically, it can be treated as an clustering problem and then converted into an optimization problem with a customized objective function. MCMC is perfect methodology for optimization problem like this. How to scale this solution for Big Data is another interesting topic. Dr Luo has designed and implemented a large-scale disambiguation system to identify and disambiguate multi-sense skills, using Markov Chain Monte Carlo (MCMC). More details can be found in his paper (Macau: Large-Scale Skill Sense Disambiguation in the Online Recruitment Domain). Restricted Boltzmann Machine (RBM) and Recommendation Engines: RBM is one of the most famous Deep Learning (DL) algorithms which can be used in commercial recommendation systems. RBM was inspired by the concepts from Physics (Energy, Boltzmann distribution, partition function) and I am interested in: (1) using other distributions and partition functions from Statistical Mechanics to replace Boltzmann distribution, for building new recommendation algorithms; (2) using MCMC to replace contrastive divergence or gradient descent as an optimization method to train RBM. Contrastive divergence is used to train RBM for recommendation engines. Contrastive divergence is "approximate" method of gradient descent, and MCMC can be leveraged as an "exact" methodology for optimization in RBM.

Dr. Zhifeng He

Job Titles:

Data Scientist at ADP
Industrial Research Scientist / Payroll Innovative, ADP

Dr. Zhifeng He is a Data Scientist at ADP. Previously he was a Data Scientist at Travelport. He got his PhD in Electrical and Computer Engineering from Auburn University, Auburn, AL. He is interested in machine learning, deep learning, optimization, distributed computing, and their applications in wireless networks. The research projects Dr. He has been working on includes: Distributed computing platform for large-scale optimization problems in wireless networks; in traditional wireless networks, optimization problems such as resource allocation, data transfer scheduling are solved at the Base Station in a centralized manner, and decisions are then sent to the mobile users. This infrastructure introduces a heavy computational burden to the Base Station since all the computations are done at the Base Station, which often has limited computational power, and it takes long time for the Base Station to solve the problem; as the network size increases, the size of the optimization problem becomes too big for the Base Station to solve. How to design distributed optimization mechanisms where the computational burden at the Base Station can be shared by the mobile users to accelerate the solving of large-scale optimization problems, is an interesting topic. Dr. He proposed a Column Generation algorithm-based distributed computing system to accelerate the computation to solve large-scale optimization problems efficiently in wireless networks. The optimization problem is solved collaboratively by Base Station and worker nodes mobile users, which greatly relieves the computational burden of the Base Station and thus significantly accelerates the computational process. Besides, Dr. He also proposed a novel communication mechanism between Base Station and mobile users to reduce communication time and power consumption for the distributed computing system by minimizing the exchanged information volume. This project is an ongoing work of this paper: Zhifeng He, Shiwen Mao, and Sastry Kompella, "A decomposition approach to quality of service driven multi-user video streaming in cellular cognitive radio networks," IEEE Transactions on Wireless Communications, vol.15, no.1, pp.728-739, Jan. 2016. DOI: 10.1109/TWC.2015.2477509. Online learning time series prediction and paralleled computing library implementation: time series prediction is widely applied in different areas such as demand forecast and stock forecast. Traditional statistical time series prediction algorithms such as ARIMA is computational demanding in that each time when a new data comes in, due to the uncertainty of the time series data, it needs to recalculate all the model parameters and coefficients to get a new model, in order the capture the new pattern of the time series data. However, recalculating all the model parameters and coefficients is computational expensive, especially when there are billions of different time series data to predict. Dr. He proposed an online learning algorithm to fine tune the model at each time slot by minimizing the difference between predicted value and the observed true value at each time slot using Gradient Descent. The proposed method is shown to be much faster than recalculating all model parameters and coefficients. Besides, Dr. He also implemented the proposed algorithm using CUDA which is a paralleled computing platform, so that the proposed algorithm can solve millions of different time series prediction problems in real time, which is the first time among all existing libraries.