We developed new efficient algorithms to learn local structures in the graphical models, in terms of Markov Blankets.
These local structures provide an efficient approach of structure learning.
Specifically, my work involves:
Discovered a new coexistence property between possible false positive parent-child set and the spouses in a Markov Blanket.
Developed new efficient constraint-based and score-based algorithms that improve the state-of-the-art.
We apply the Markov Blanket algorithms to feature selection. By considering structured information among variables,
this approach demonstrates several optimalities for feature selection.
Specifically, my work involves:
Proved several optimalities of structured feature selection using Markov Blanket.
Developed algorithms for large-scale and hierarchical feature selection.
We developed a new algorithms to efficiently learn the direct causes and effects of a target variable by using Markov Blankets.
The proposed local causal discovery algorithm drastically improves the efficiency from the state-of-the-art.
Specifically, my work involves:
Discovered that independence relationship changes among neighbours of the target can help detect direct causes and effects.
Designed a new algorithm for local causal discovery for direct causes and effects.
Robust Constraint-based Causal Discovery Under Insufficient Data
We propose Bayesian-augmented frequentist independence tests to improve the performance of constraint-based causal discovery methods under insuffcient data. We apply proposed independence tests to constraint-based causal discovery methods and evaluate the performance on benchmark datasets with insuffcient samples. Experiments show signifcant performance improvement in terms of both accuracy and effciency over SOTA methods.
Specifically,
We introduce a Bayesian method to estimate mutual information (MI), based on which we propose a robust MI based independence test.
We consider the Bayesian estimation of hypothesis likelihood and incorporate it into a well-defned statistical test, resulting in a robust statistical testing based independence test.
We investigated the Markov Blanket discovery under the presence of latent variables. By finding these local latent variables,
we can discover the previously unknown factors, get a more compact model with better generalization, and improve the classification results.
Specifically, my work involves:
developed an algorithm to discover the latent variables local to one target variable.
Human Body Tracking and Gesture Recognition
Fall 2011 ~ Fall 2012, NSF Triple Helix Graduate Fellowship, Rensselaer Polytechnic Institute, Troy, NY
Developed a human gesture recognition system with Microsoft Kinect camera, using dynamic graphical models.
Developed a human pose estimation framework from depth image sequences to recognize upper body parts.
Designed and co-developed two education software with the Microsoft Kinect camera with computer vision and machine learning algorithms,
providing middle school students with additional STEM learning tools.
- Tian Gao, and Qiang Ji. "Efficient Markov Blanket Discovery and Its Applications",
IEEE Transactions on Cybernetics. To Appear. 2016
- Tian Gao, and Qiang Ji. "Efficient Score-Based Markov Blanket Discovery"
Int. J. Approx. Reason., 2016. Submitted.
Conferences
- Tian Gao, and Qiang Ji. "Constrained Local Latent Variable Discovery",
International Joint Conference on Artificial Intelligence (IJCAI), 2016. Accepted.
- Tian Gao, and Qiang Ji. "Local Causal Discovery for Direct Causes and Effects",
Neural Information Processing Systems (NIPS), 2015.
- Tian Gao, Ziheng Wang, and Qiang Ji. "Structured Feature Selection",
International Conference on Computer Vision (ICCV), 2015.
- Ziheng Wang, Tian Gao, and Qiang Ji. "Learning with Hidden Information Using a Max-Margin Latent Variable Model".
22nd International Conference on Patter Recognition (ICPR), 1389-1394. IEEE, 2014.