Improving the Predictive Power of Business Performance Measurement Systems by Constructed Data Quality Features? Five Cases

• Published in: Advances in Data Mining: Applications and Theoretical Aspects pp. 3 - 16
• Main Author: Vattulainen, Markus
• Format: Book Chapter
• Language: English
• Published: Cham Springer International Publishing 2015
• Series: Lecture Notes in Computer Science
• Subjects: Business performance measurement system; Data quality; Feature construction; Predictive classification; Preprocessing
• ISBN: 9783319209098; 3319209094
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-319-20910-4_1

Predictive power is an important objective for current business performance measurement systems and it is based on metrics design, collection and preprocessing of data and predictive modeling. A promising but less studied preprocessing activity is to construct additional features that can be interpreted to express the quality of data and thus provide predictive models not only data points but also their quality characteristics. The research problem addressed in this study is: can we improve the predictive power of business performance measurement systems by constructing additional data quality features? Unsupervised, supervised and domain knowledge approaches were used to operationalize eight features based on elementary data quality dimensions. In the case studies five corporate datasets Toyota Material Handling Finland, Innolink group, 3StepIt, Papua Merchandising and Lempesti constructed data quality features performed better than minimally processed data sets in 29/38 and equally in 9/38 tests. Comparison to a competing method of preprocessing combinations with the first two datasets showed that constructed features had slightly lower prediction performance, but they were clearly better in execution time and easiness of use. Additionally, constructed data quality features helped to visually explore high dimensional data quality patterns. Further research is needed to expand the range of constructed features and to map the findings systematically to data quality concepts and practices.