Explaining regression prediction with contributions of input variables

Julian Klauser (2010) Explaining regression prediction with contributions of input variables. EngD thesis.

PDF Download (6Mb)

Abstract

The explanation of predictions, made by regression algorithms, is a vey important aspect of the discovery of patterns in data. In this thesis we extend an existing method for explanation in classification, developed by Štrumbelj and Kononenko, to regression. The base of the method comes from game theory and with the help of sampling, solves the problem of exponential time growth. We presented the basic concept of the method and after its successful implementation, we tested it with various regression algorithms on artificial domains. We analyzed the performance and efficiency of the presented method and compared algorithms according to the success of the prediction and the success of the explanation. Besides the existing visualisation of the prediction's explanation we created a new way of visualising the explanation models. We have also shown that our method can be extended, so that it is able to cleave to the context with partial compliance to attribute value distribution.

Actions (login required)