Abstract
This paper presents Interval Prediction Tree INPRET algorithm for interval prediction of numerical target variables from temporal mean-variance aggregated data. The proposed algorithm allows to process mean-variance aggregated multivariate temporal data and to identify outliers in training data instances. The proposed algorithm enables, on the one hand, to utilize predictive feature information obtained from mean and variance of temporally aggregated instances, and on the other hand, to achieve a considerable reduction in the depth of the induced prediction tree by using interval prediction tree leaves. As shown by our empirical evaluation of aircraft maintenance real world multi-sensor data set, in terms of the prediction tree size and the root mean square error, the proposed algorithm provides better integration between accuracy and performance than existing regression tree models.