Contrastive explanation of the output of machine learning models

Abstract

In this research we aim to automatically generate an explanation of decisions made by machine learning models. To be able to do this we adapted the explanation based model by Feelders and Daniels. We describe the building blocks of the model and consider different options for determining the required reference object. In this thesis we calculate the required distances with either Gower's distance or the simplex method. For the type of object reference we use either the closest object with the desired classification or the medoid object of the desired classification. We test the proposed algorithm with a questionnaire that tested the quality of the explanation and parts there of. We found that the medoid reference type was significantly better received by respondents than the closest reference type. In the ordinal logistic regression model we found a significant negative effect of the number of errors people made in the subject knowledge questions on the perceived explanation quality. We were unable to find any significant results for the other factors, and we found no significant effect on which distance function performs better, or if adding the counter-acting to the contributing causes had an effect on the overall perception of the explanation. As the number of participants to the empirical study was rather small we opted to go for a more exploratory approach to find factors that could be interesting to investigate in further studies. Due to this reason future research with this method is recommended as we were unable to get a definitive conclusion for our developed model.