Bayesian Network with Residual Correction Mechanism

Abstract

A major issue in spatio-temporal (ST) prediction of any variable is the unavailability of the data on influencing factors. This happens, because it is not always known properly which variable influences which other. In that case, modeling of spatio-temporal inter-relationships using graphical model (like Bayesian network) becomes a challenging task due to the lack of appropriate influencing nodes in the dependency graph . In this chapter, we introduce a novel architecture of BN analysis with incorporated residual-correction mechanism  (BNRC). The embedded residual-correction mechanism in BNRC helps to compensate for the unavailable variables in the causal dependency graph  of Bayesian network , and thereby assists in improving the accuracy when adopted in a prediction model. The performance of BNRC has been evaluated in comparison with a number of conventional statistical and state-of-the-art space-time prediction models, with respect to case studies on climatological and hydrological time series prediction . Experimental result demonstrates effectiveness of BNRC in spatial time series prediction under the paucity of influencing variables.