Bayesian classification results using data containing missing class labels

Abstract

95Abmuct-In this paper, the Mean-Field Bayesian Data Reduction Algorithm has been extended to adaptively train on data containing missing values. In the basic data model for this algorithm each feature vector of a given class contains a class-labeling feature. That is, while training on a data set the Mean-Field Bayesian Data Reduction Algorithm labels the unlabeled adapted data, while simultaneously determining those features that provide best classification performance. This has the benefit of improving performance by reducing the data to mitigate the effects of the curse of dimensionality.In the results shown, the Mean-Field Bayesian Data Reduction Algorithm is used to classify real-life data sets found at the UCI Repository. The emphasis in the results is on demonstrating performance for situations in which increasing numbers of training data are lacking any class labels. This is useful in illustrating the effect that removing class labels has on the overall probability of error in correct classification, as compared to the certain labeled case. In general, it is shown that the methods developed here are can significantly improve performance when the proportion of uncertain-labeled training data is very large.

Two‐group classification using the Bayesian data reduction algorithm

“…The Bayesian data reduction algorithm (BDRA) is a relatively new technique developed by Lynch and Willet [8][9][10][11]. The BDRA algorithm, being based on Bayes' rule, shares some theoretical foundations with traditional techniques.…”