Explanation and Reliability of Individual Predictions
Igor Kononenko, Erik Štrumbelj, Zoran Bosnić, Darko Pevec, Matjaž Kukar, Marko Robnik‐Šikonja · 2013
Classification and regression models, either automatically generated from data by machine learning algorithms, or manually encoded with the help of domain experts, are daily used to predict the labels of new instances. Each such individual prediction, in order to be accepted/trusted by users, should be accompanied by an explanation of the prediction as well as by an estimate of its reliability. We have recently developed a general methodology for explaining individual predictions as well as for estimating their reliability. Both, explanation and reliability estimation are general techniques, independent of the underlying model and provide on-line (effective and efficient) support to the users of prediction models. Povzetek: Razvita je metodologija za razlago napovedi s pripadajočo verjetnostjo pri klasifikacijskih in regresijskih modelov strojnega učenja.