Handling missing features in maximum margin Bayesian network classifiers

Abstract

The Comprehensive Nuclear-Test-Ban Treaty Organization (CTBTO) records hydroacoustic data to detect nuclear explosions1. This enables verification of the Comprehensive Nuclear-Test-Ban Treaty once it has entered into force. The detection can be considered as a classification problem discriminating noise-like, earthquake-caused and explosion-like data. Classification of the recorded data is challenging because it suffers from large amounts of missing features. While the classification performance of support vector machines has been evaluated, no such results for Bayesian network classifiers are available. We provide these results using classifiers with generatively and discriminatively optimized parameters and employing different imputation methods. In case of discriminatively optimized parameters, Bayesian network classifiers slightly outperform support vector machines. For optimizing the parameters discriminatively, we extend the formulation of maximum margin Bayesian network classifiers to missing features and latent variables. The advantage of these classifiers over classifiers with generatively optimized parameters is demonstrated in experiments.