Learnability of Dynamic Bayesian Networks from Time Series Microarray Data

Abstract

Dynamic Bayesian networks (DBNs) are becoming widely used to learn gene regulatory networks from time series microarray data. Careful experimental design is required for data generation, because of the high cost of running each microarray experiment. This paper presents a theoretical analysis of learning DBNs without hidden variables from time series data. The analysis reveals, among other lessons, that under a reasonable set of assumptions a fixed budget is better spent on many short time series than on a few long time series. Keywords: dynamic Bayesian networks, gene expression microarrays, gene regulatory networks, PAC-learnability, time series data