Investigating the effectiveness of conditional classification: an application to manufacturing scheduling

This paper examines the problem of multidimensional classification, an automated learning process where "rules" are to be inferred on separate but related aspects of a problem, using identical or overlapping data sets. A general framework describing the various types of multidimensional classification is provided. The paper specifically concentrates on conditional classification, wherein the order of classification is based on domain semantics. Drawing from concept learning and information theory, algorithms are presented for acquiring tree-structured knowledge from available data. An application to manufacturing scheduling is presented. Results indicate that conditional classification may provide some ability to better interpret related decisions in automated manufacturing contexts. Further work is necessary to ascertain if the approach is robust, particularly on more complex decisions, larger data sets, and noisy data. >