Pattern Classification in Context

Oregon Health & Science University, Portland OR

Investigators

Abstract

In the real world, the identity of an object is often ambiguous due to noise or lack of information. This ambiguity can sometimes be reduced by utilizing extra information, referred to as context, provided by surrounding objects. Unfortunately, existing approaches that utilize context collapse when the number of context-bearing objects is large. Furthermore, there has been no attempt to separate relevant context from irrelevant context. In this project PI will develop algorithms that can reliably identify relevant context and then exploit it in a decision making process for pattern classification that is computationally tractable. The approach employs a Bayesian framework that incorporates "partial" context represented as the "derivative" of the identities of surrounding objects. The criteria for evaluation of context relevancy include information-theoretic measures such as mutual information and "expected relative entropy", as well as complete factorization of Bayesian networks. The development of the algorithms was initially motivated by and will be ultimately tested on two related yet different medical diagnostic applications: white blood cell differentiation and microscopic urinalysis image classification. The resulting classification will provide diagnostic information about real patients. Although the testbed applications relate to medical diagnosis, the results of this research will be general and can find an abundance of uses in other areas, so this work will impact the broader scope of knowledge discovery, data mining, machine learning, and pattern recognition where context plays a role.