Efficient Reasoning upon Fusion of Many Data Sources

Abstract

Bayesian Knowledge Bases (BKB), a graphical model for representing structured probabilistic information, allow for efficient fusion of knowledge from multiple sources. Past research has focused on knowledge fusion situations that only involve a limited number of sources. In this work, we extend the BKB fusion research by exploring the effect of fusing information from many sources. This extension quickly yields a reasoning bottleneck that we overcome by leveraging a representation modification algorithm.  With this algorithm, we show how reasoning complexity upon fusion of many sources can be significantly reduced while maintaining the underlying probability semantics of all sources. This develops a means for BKBs to be used in various data fusion problems, allowing previously intractable problems to be studied. We further illustrate our solution empirically using two simulated problems as well as practically through survival time analysis of breast cancer data taken from The Cancer Genome Atlas (TCGA) Program.