Transactions on Machine Learning and Data Mining (ISSN: 1865-6781)
Volume 11 - Number 1 - July 2018 - Pages 27-41
Parsimonious Modeling for Binary Classification of Quality in a High Conformance Manufacturing Environment
Carlos A. Escobar Diaz¹,&178, Ruben Morales-Menendez²
¹General Motors, Research and Development, Warren MI 48092, USA,
²Tecnologico de Monterrey, Monterrey, NL. 64849, Mexico,
Abstract
The world of big data is changing dramatically; in the domain of data mining, machine learning and pattern recognition, the feature access has grown from tens to hundreds or even thousands. This trend presents enormous challenges, specially for classification problems. In manufacturing, classification of quality is one of the most important applications; however, feature explosion, combined with high conformance production rates are two of the most important challenges for big data initiatives. Empirical evidence shows that discarding irrelevant or redundant features improves prediction, helps in understanding the system, reduces running time requirements, and reduces the effect of dimensionality. In this paper, the Hybrid Correlation- and Ranking-based (HCR) and ReliefF filter feature elimination algorithms are presented as a wrapper method, which uses the Naive Bayes as the learning algorithm. To boost parsimony, the algorithms are combined with the Penalized Maximum Probability of Correct Decision – a model selection criterion – to develop a Hybrid Feature Selection and Pattern Recognition framework aimed at rare quality event detection. A ?exible approach that can be widely applied to various machine learning algorithms.
Keywords: Quality control, Manufacturing systems, Feature elimination algorithm, Model selection criterion, Unbalanced binary data, Defect detection
Download Paper
(379 KB)