Robustness of Predictive Data Mining Methods under the Presence of Measurement Errors in the Context of Production Processes

Abstract

One of the main objectives of data analysis in industrial contexts is prediction, that is, to identify a function that allows predicting the value of a response from the values of other variables considered as potential predictors of this outcome. The large volumes of data that current technology allows to generate and store have made it necessary to develop methods of analysis alternative to the traditional ones to achieve this objective, which allow mainly to process these large amounts of information and to predict the response in real time. Enclosed under the name of Data Mining, many of these new methods are based on automatic algorithms mostly originated in the computer field. However, the quality of the information that feeds these procedures remains a key factor in ensuring the reliability of the results. With this premise, in this work we study the effect that the presence of faults in the measurement devices that originate the information to be analyzed, can cause on the predictive ability of one of the predictive methods of data mining, the decision trees. The results are compared with those obtained using one of the traditional statistical techniques: multiple linear regression. The results obtained indicate that the effect of measurement related errors on the predictive ability of decision trees, compared to traditional regression models, depends on the nature of the measurement error.