My Personal Website at University of Michigan

“Changing the world, one site at a time...”

Adaptive Design for developing personalized treatment

K. Deng, J. Pineau and S.A.Murphy (2011). Active Learning for Developing Personalized Treatment. Proceedings of the Twenty-Seventh Conference Annual Conference on Uncertainty in Artificial Intelligence (UAI-11). AUAI Press 161-8. Presentation slides.

K. Deng, J. Pineau and S.A.Murphy (2011). Active Learning for Personalizing Treatment. Adaptive Dynamic Programming And Reinforcement Learning (ADPRL), 2011 IEEE Symposium on. 11-15 April 2011, pages 32-39. Presentation slides.

Budgeted Learning
Some machine learning applications have training data where the labels are available but the attributes describing the examples must be purchased. This problem is, in a sense, a dual to active learning. Budgeted learning algorithms attempt to intelligently choose the attributes to purchase in order to learn well with as few purchased attributes as possible.

Kun Deng, Chris Bourke, Stephen Scott, Julie Sunderman, and Yaling Zheng. Bandit-Based Algorithms for Budgeted Learning. In Proceedings of the Seventh IEEE International Conference on Data Mining. October 2007, pages 463-468.
Abstract
paper (849Kb PDF)

Kun Deng, Chris Bourke, Stephen Scott, Julie Sunderman, and Yaling Zheng. Bandit-Based Algorithms for Budgeted Learning, accepted by Machine Learning Journal. Final version coming soon!


Active Learning
Some machine learning applications have abundant amounts of unlabeled data with an oracle that is capable of labeling a relatively small number of these examples to be used in supervised training. Active learning algorithms attempt to intelligently choose the examples to label in order to learn well with as few labeled examples as possible.


Yaling Zheng, Stephen Scott, and Kun Deng. Active Learning from Multiple Noisy Labelers with Varied Costs. In Proceedings of the Tenth IEEE International Conference on Data Mining, pages 639–648. December 2010.
paper (223Kb PDF)



Matt Culver, Deng Kun, and Stephen Scott. Active Learning to Maximize Area Under the ROC Curve. In Proceedings of the Sixth IEEE International Conference on Data Mining. December 2006, pages 149-158.
Abstract
paper (146Kb PDF)

Thomas Osugi, Deng Kun, and Stephen Scott. Balancing Exploration and Exploitation: A New Algorithm for Active Machine Learning. In Proceedings of the Fifth IEEE International Conference on Data Mining, pages 330-337. November 2005.
Abstract
paper (340Kb PDF)

Receiver Operating Characteristic (ROC) Analysis
Receiver Operating Characteristic (ROC) analysis is an alternative means of measuring classifier peroformance. We study mechanisms of adjusting multi-class classifiers to improve performance (with respect to its ROC hypersurface) when faced with nonuniform misclassification costs.

Kun Deng, Chris Bourke, Stephen Scott, and N. V. Vinodchandran. New Algorithms for Optimizing Multi-Class Classifiers via ROC Surfaces. In Proceedings of The Third Workshop on ROC Analysis in Machine Learning, pages 17-24, June 2006.
Abstract
264Kb PDF

Chris Bourke, Kun Deng, Stephen D. Scott, Robert Schapire, and N. V. Vinodchandran. On reoptimizing multi-class classifiers. Machine Learning, 71(2–3):219–242; doi:10.1007/s10994-008-5056-8, 2008.
Abstract
On-line version


Medical Imaging Applications

Hong Wu, Kun Deng, Jianming Liang, Machine learning based automatic detection of pulmonary trunk. In SPIE Conference for Computer Aided Diagnosis.
Abstract:
Pulmonary embolism is a serious condition causing sudden death in about one-third of the cases. Computed tomographic pulmonary angiography has become the diagnostic standard for PE. Several approaches for computer aided diagnosis of PE in CTPA have been proposed, but they are not effective in detecting central PEs, handling the false positives from the pulmonary veins and adapting to suboptimal contrast conditions. Overcoming these deficiencies demands highly efficient and accurate identification of the pulmonary trunk, therefore, this paper presents a machine learning based approach, achieving a nearly 100% accuracy tested on a large number of cases.