Scalable Learning of Bayesian Network Classifiers

摘要

Ever increasing data quantity makes ever more urgent the needfor highly scalable learners that have good classificationperformance. Therefore, an out-of-core learner with excellenttime and space complexity, along with high expressivity (thatis, capacity to learn very complex multivariate probabilitydistributions) is extremely desirable. This paper presents sucha learner. We propose an extension to the $k$-dependenceBayesian classifier (KDB) that discriminatively selects a sub-model of a full KDB classifier. It requires only one additionalpass through the training data, making it a three-pass learner.Our extensive experimental evaluation on $16$ large data setsreveals that this out-of-core algorithm achieves competitiveclassification performance, and substantially better trainingand classification time than state-of-the-art in-core learnerssuch as random forest and linear and non-linear logisticregression. color="gray">