Projects

• Learning hierarchical Bayesian models with nonparametric priors for user/customer profiles
  

  We construct hierarchical Bayesian models to model user preference. Nonparametric priors are used in such models. This provides great flexibility in modeling the true distribution of user preference which leads to better predictions.
  

  (Current project)




• Adaptive and proactive cyber-defense with Gaussian process learning
  

  Intrusion attackers have different preference on the attack tools they use and the networks they target. We model individual attacker preference by a preference function and the distribution of such preference functions among the attacker population by a Gaussian process prior. The Bayesian model is learned from shared cyber-alert reports and then used for predicting future attacks.
  

  (Oct, 06 -- Oct, 07, Collaborators: Phil Porras, Johannes Ullrich)




• Coclustering by similarity refinement
  

  We proposed a novel framework for coclustering related objects such as documents and the words contained in these documents. In this framework, some approximate clustering structure on one type of objects (e.g. terms) is used to refine the similarity between the objects of the other type (e.g. documents). An alternating refinement process leads to better clustering of both types of objects.
  

  (Jan, 07 -- Jun, 07)




• Scalable mining of correlations
  

  We designed and implemented a scalable algorithm for mining correlations among objects in extremely large data sets using locality sensitive hashing. The algorithm outperforms the next best algorithm by several orders of magnitude.
  

  (Jan, 06 -- Jun, 06, Collaborator: Joan Feigenbaum)




• BGP update anomaly detection by wavelet analysis
  

  We apply wavelet analysis to extract features from BGP update sequences. The features are then used by a nearest neighbor learning process to identify outliers/anomalies in BGP updates.
  

  (Sep, 04 -- May, 05, Collaborators: Joan Feigenbaum, Jennifer Rexford)




• Efficient graph and geometry algorithms for data streams
  

  We designed data-stream specific algorithms for geometric and graph problems such as diameter and shortest-path distances. We also proved streaming-space lower bounds for these problems.
  

  (2002 -- 2004, Supervisor: Joan Feigenbaum)




• Numerical simulation of neuron activities and networked neurons
  

  We design and implement computational models to simulate how human attention affects their vision process.
  

  (1999, Supervisor: Larry Abbott)