Influence of Domain and Model Properties on the Reliability Estimates' Performance

In machine learning, the reliability estimates for individual predictions provide more information about individual prediction error than the average accuracy of predictive model (e.g. relative mean squared error). Such reliability estimates may represent decisive information in the risk-sensitive applications of machine learning (e.g. medicine, engineering, and business), where they enable the users to distinguish between more and less reliable predictions. In the authors’ previous work they proposed eight reliability estimates for individual examples in regression and evaluated their performance. The results showed that the performance of each estimate strongly varies depending on the domain and regression model properties. In this paper they empirically analyze the dependence of reliability estimates’ performance on the data set and model properties. They present the results which show that the reliability estimates perform better when used with more accurate regression models, in domains with greater number of examples and in domains with less noisy data.