| Abstract: |  The thesis analyses methods for assessing usability of different regression models. One type of methods are heuristically inspired, designed so that they correlate with prediction error. Another type of methods consists in obtaining prediction intervals using either frequentist statistical methods or transductive inference. Based on the analysis, two modification of existing heuristic methods were proposed. Both the methods types are compared on four kinds of parametric regression models using two real-world datasets. The analysis of results showed particular cases where the correlation of the heuristic methods' output with prediction error is relatively high, which was then also showed by obtained Pearson coefficient of linear dependence on an independent set of observations. The presented results also showed that the heuristic method output is strongly dependent on used regression model and the dataset. However, one of the proposed method embodies relatively high correlation in case of one dataset and all the regression models.
|
|---|