Ikea effect in rule-based machine learning models

This thesis tries to evaluate the effect of a selected cognitive bias on the interpretability of rule-based machine learning models. This thesis focuses on the "IKEA effect" bias, according to which people assign a higher value to things that they created or partially created. Applied to machine leaning models, the hypothesis investigated in this thesis is that models that people had a chance to edit or improve will be more plausible and comprehensible than those where such opportunity was not provided. To test this hypothesis, the crowdsourcing experiment was carried out. In these experiments, participants had to evaluate the plausibility of rules automatically generated by the Apriori algorithm. One group of participants was additionally invited to use a rule editor to improve the rules. In this preliminary study, no difference in the elicited comprehensibility and plausibility ratings were observed. The data for statistical analysis provided by the crowdsourcing platform accounts for 120 answers with an equal sample size distribution between control and experimental groups processed through the selection procedure. Applying t-test and the non-parametric test shows the p-values that did not allow to reject the null hypothesis. However, the test of effect size shows a small effect for both aspects of interpretability. The advice for future works is to create more experiments using various rule sets from different topics and involve more people to achieve higher accuracy in statistics to obtain a more reliable result.