Estimation of Censored Data Water Quality Values Using Decomposable Markov Networks

Abstract

While application of probabilistic inference modeling to large and complex datasets has been limited both as a result of computational difficulties, and implicit/explicit assumptions of normality and lognormality, an alternative is developed herein, based on advancements in graphical modeling using decomposable Markov networks (DMNs). Uncertainties in estimates for censored and/or missing data, are reduced by quantifying dependencies among quality attributes using DMNs. The dependence structure is modeled by a DMN, and established using training data. The improvement from learning DMNs employing the training data is demonstrated using water quality information from water distribution systems. The approach provides a general alternative to traditional techniques for estimating values for censored data.

Keywords: Censored data, inductive learning, knowledge acquisition, Markov networks, model mining, water quality