Project: Kernel Learning For Fusing Uncertain Information From Multiple Heterogeneous Sources 

(ONR N00014-09-1-0663)

PI: Anil K. Jain, Department of CSE, Michigan State University

Co-PI: Rong Jin and Pang-Ning Tan, Department of CSE, Michigan State University

Abstract:

This project will develop a computational framework and efficient algorithms for combining uncertain information from different sources, both quantitative and qualitative, in a unified framework. In particular, we propose a kernel based framework for information fusion that extracts and combines noisy and uncertain information from multiple heterogeneous sources. The key idea is to represent data observations and side information of different types and in different formats uniformly by a kernel matrix. The framework consists of two major components: (1) Kernel learning. Learn a kernel data representation from noisy observation and uncertain side information. This component explicitly addresses the challenge in computing kernel matrices given the noisy observations and uncertainties in side information. (2) Multiple kernel combination. Find the optimal combination of kernels via a maximum margin framework. This component explicitly addresses the challenges in combining multiple kernel matrices in a nonlinear fashion and constructing an effective and reliable decision function given the noisy kernel data representations and uncertain side information.The proposed framework will be applied to a real-world application, namely, information fusion for dynamic social network analysis.

Students

1. Tianbao Yang

Conference:

1.  Steven Hoi, and Rong Jin, Online Multiple Kernel Learning, submitted to Journal of Machine Learning Research, 2009