Improving regression predictions using individual point reliability estimates based on critical error scenarios

We present a method to improve computed regression prediction values for unseen data. It aims at obtaining more accurate results by adjusting the calculated predictions instead of by constructing a different regression model. As a result, it can be helpful to improve the prediction of a specific observation provided by an existing benchmark regression model or predictor system. The proposed methodology uses individual point reliability estimates that indicate if a single regression prediction is likely to produce an error considered critical by the user of the regression. We tested the method in two sets of experiments, one using synthetically produced data, and the other using data from the public data repository UCI Machine Learning. The experiments with synthetic data were performed to verify the efficiency of the method under controlled situations. In this case, the method produced superior results improving predictions for cleaner data with progressive worsening with the increase of the noise level. Experiments with ten databases from the UCI data repository were executed to investigate the applicability of the methodology using real world data. The method was able to correctly adjust regressions prediction values in experiments with all the ten databases, achieving statistically significant improvement in eight of them.