Home // DATA ANALYTICS 2021, The Tenth International Conference on Data Analytics // View article

Evaluation of Filter Methods for Feature Selection by Using Real Manufacturing Data

Authors:
Alexander Gerling
Holger Ziekow
Ulf Schreier
Christian Seiffer
Andreas Hess
Djaffar Abdeslam

Keywords: Filter-Based Feature Selection methods; Machine Learning; Cost based Metric; Production; Production data

Abstract:
The importance of Machine Learning (ML) in the domain of manufacturing has been increasing in recent years. Especially, ML techniques are used to predict and explain errors in the production. One challenge of using ML in this domain is to deal with the often-high number of features in the datasets. However, a product defect can in many cases be traced back to a few relevant characteristics. In this paper, we investigate methods for finding reduced feature sets in the context of manufacturing. Here, the feature reduction promises two key advantages. One improvement is the prediction quality of the ML model. The second advantage concerns the explainability of a product error. With a reduction of features from the original dataset, we also reduce the search space for the product error origin. We investigate three different filter methods for feature selection based on 25 real manufacturing datasets, which are highly unbalanced. We describe the implementation of these and test them in three experimental approaches. Furthermore, we optimize the feature selection using a cost-based metric. Optimizing on the basis of the costbased metric is shown to be in several cases more useful for reducing the number of features than well-established and frequently used classification metrics. In various experiments, we were able to improve the result and simultaneously reduce the number of features with our cost-based metric.

Pages: 82 to 91

Copyright: Copyright (c) IARIA, 2021

Publication date: October 3, 2021

Published in: conference

ISSN: 2308-4464

ISBN: 978-1-61208-891-4

Location: Barcelona, Spain

Dates: from October 3, 2021 to October 7, 2021