Novel ‘hybrid’ classification method employing Bayesian networks

Abstract

Standard statistical discriminant analysis techniques inherently make assumptions about underlying class structures in data, limiting their validity and effectiveness. Other classification methods, such as soft independent modeling of class analogy (SIMCA) or artificial neural networks, replace the disadvantage of making such assumptions with an equally impeding lack of interpretability. The intention of this work was to formulate a classification scheme that avoids these and other obstacles. A new classification technique has been designed that combines the recursive partitioning feature of tree‐based classifiers (e.g. classification and regression trees (CART)) with the probabilistic reasoning of Bayesian networks. The proposed hybrid approach benefits from all the advantages of tree‐based classifiers and Bayesian networks without experiencing the usual limitations associated with these methods individually. The resulting classifier outperformed several standard methods and has the added benefits of being both statistically and semantically justifiable. Copyright © 1999 John Wiley & Sons, Ltd.

Metabonomic Assessment of Physiological Disruptions Using 1H−13C HMBC-NMR Spectroscopy Combined with Pattern Recognition Procedures Performed on Filtered Variables

“…In fact, within-group covariance matrices are considered as identity in LDA 53 whereas PLS do not formulate such a hypothesis . This lack of interpretability has been already pointed out and may be related to the nonlinear modeling involved by PLS . In terms of model parsimony, LDA summarizes between-group variance (LD1 44%, LD2 35.2%) in a more efficient way than PLS-DA (t[1] 44.5%, t[2] 9.7%) on the first two discriminant axes.…”

Hybrid Bayesian networks: making the hybrid Bayesian classifier robust to missing training data

“…The work reported here focuses on augmenting the hybrid classifier previously developed in this laboratory [2] to produce a classifier that would be more robust to missing values in the data used to develop the classification model.…”

“…The hybrid Bayesian classifier [2] referrred to above combines the feature-selective power of classification and regression trees (CART) with the probabilistic nature of Bayesian systems. This hybrid system was created by using CART [6±9] to produce highly discriminating discrete variables that could in turn be used to build a Bayesian classifier [10±12].…”

“…The thyroid dataset provides another simple example of the hybrid Bayesian network and was the dataset used to demonstrate the original hybrid Bayesian classifier [2]. The goal in the analysis of this dataset is to determine the medical status of the thyroid given data from several different blood tests.…”

Discriminant Analysis and Class Modelling of Spectroscopic Data, E. K. Kemsley, Wiley, Chichester, 1998, ISBN 0‐471‐97830‐2, xii + 179 pp. £ 500·00 (including software).

“…Finally, I miss, at least, one sentence that informs the reader that the methods described in the book are the most simplest tools of classification, and some hints for further reading [4]. Inquisitive users should read some more recent articles [5,6] as well.…”