Enhanced Data Modelling Approach with Interval Estimation

Abstract

This paper deals with regression tasks on real-valued numerical data attributes. A special data transformation formulated in our earlier work is combined with a new and enhanced weighting strategy in order to improve prediction accuracy. The proposed data modelling approach offers several advantages: it does not depend on the particular regression model used and it enables the analyst to calculate tolerance interval estimates as well as the probability that the target attribute exceeds arbitrary predefined thresholds. We tested our approach on three real-world datasets. In all the three cases it reliably improved and stabilized the prediction accuracy (measured by the average root mean squared error for each dataset) as well as the quality of tolerance interval estimates.