文档库

最新最全的文档下载
当前位置:文档库 > DEMPSTER-SHAFER THEORY

DEMPSTER-SHAFER THEORY

Proceedings of the 3rd INFORMS Workshop on Data Mining and Health Informatics (DM-HI 2008)

J. Li, D. Aleman, R. Sikora, eds.

DEMPSTER-SHAFER THEORY BASED SIMULTANEOUS LEARNING OF MULTIPLE BAYESIAN NETWORKS

Shuai Huang

Department of Industrial Engineering

Arizona State University

Tempe, AZ

Jing Li

Department of Industrial Engineering

Arizona State University

Tempe, AZ

Jinglz@asu.edu

Abstract

Bayesian networks (BNs) have been used extensively in medical diagnosis due to their capability of learning causal relationships from data. Most existing BN structure learning algorithms have focused on learning a BN for a single problem/task. This overlooks the situations when data are available for multiple related problems/tasks, such as patients in different age cohorts, or with different genetic signatures. In these situations, learning BNs for multiple related problems/tasks simultaneously may improve the learning efficiency and effectiveness. Therefore, this paper develops a new algorithm to achieve simultaneous learning of BN structures for multiple related problems/tasks, which integrates an existing BN structure learning algorithm, called K2, and the Dempster-Shafer theory for combining different sources of evidences. The developed algorithm is demonstrated using the ALARM dataset collected for studying patient care in ICU. Keywords: Bayesian network, causal inference, Dempster-Shafer theory, multitask learning

1. Introduction

A Bayesian network (BN) is a graphical model to represent dependent, independent, and even causal relationships among the variables in a multivariate system, which is widely applied to AI, medical, and manufacturing areas [1]. A BN includes a structure and parameters: the structure of a BN is a directed acyclic graph (DAG), i.e., a set of variables connected by directed arcs, directions of arcs imply causalities, and there is no path beginning and ending with the same variable; the parameters of a BN are conditional probability distributions measuring strength of the causalities.

Recent advances in computation ability have created a tremendous opportunity for designing different kinds of algorithms to learn BN structures from data [2-4,8]. However, most of these algorithms have focused on learning the BN structure for a single problem/task. This overlooks the situations when data are available for multiple related problems/tasks. In these situations, learning BN structures for multiple related problems/tasks simultaneously may improve the learning efficiency and effectiveness.

免费下载Word文档免费下载: DEMPSTER-SHAFER THEORY

(共6页)