A Hierarchical, Fuzzy Inference Approach to Data Filtration and Feature Prioritization in the Connected Manufacturing Enterprise

Abstract

The current big data landscape is one such that the technology and capability to capture and storage of data has preceded and outpaced the corresponding capability to analyze and interpret it. This has led naturally to the development of elegant and powerful algorithms for data mining, machine learning, and artificial intelligence to harness the potential of the big data environment. A competing reality, however, is that limitations exist in how and to what extent human beings can process complex information. The convergence of these realities is a tension between the technical sophistication or elegance of a solution and its transparency or interpretability by the human data scientist or decision maker. This dissertation, contextualized in the connected manufacturing enterprise, presents an original Fuzzy Approach to Feature Reduction and Prioritization (FAFRAP) approach that is designed to assist the data scientist in filtering and prioritizing data for inclusion in supervised machine learning models. A set of sequential filters reduces the initial set of independent variables, and a fuzzy inference system outputs a crisp numeric value associated with each feature to rank order and prioritize for inclusion in model training. Additionally, the fuzzy inference system outputs a descriptive label to assist in the interpretation of the feature’s usefulness with respect to the problem of interest. Model testing is performed using three publicly available datasets from an online machine learning data repository and later applied to a case study in electronic assembly manufacture. Consistency of model results is experimentally verified using Fisher’s Exact Test, and results of filtered models are compared to results obtained by the unfiltered sets of features using a proposed novel metric of performance-size ratio (PSR).