Neural networks regression inductive conformal predictor and its application to total electron content prediction

Abstract

In this paper we extend regression Neural Networks (NNs) based on the Conformal Prediction (CP) framework for accompanying predictions with reliable measures of confidence. We follow a modification of the original CP approach, called Inductive Conformal Prediction (ICP), which enables us to overcome the computational inefficiency problem of CP. Unlike the point predictions produced by conventional regression NNs the proposed approach produces predictive intervals that satisfy a given confidence level. We apply it to the problem of predicting Total Electron Content (TEC), which is an important parameter in trans-ionospheric links. Our experimental results on a dataset collected over a period of 11 years show that the resulting predictive intervals are both well-calibrated and tight enough to be useful in practice. © 2010 Springer-Verlag Berlin Heidelberg.