Reduction of search space for constructive induction using explanationsVOUK, BOŠTJAN (Avtor) Guid, Matej (Mentor) explainable artificial intelligenceexplanations of individual predictionsfeature constructionFeature construction can add to the comprehensibility and performance of machine learning models. However, unfortunately it typically requires an exhaustive search in the attribute space or time-consuming human efforts to generate meaningful features. In the dissertation, this challenge is addressed and a novel heuristic approach for reducing the search space based on the aggregation of instance-based explanations of predictive models is proposed. The dissertation, presents an efficient method for constructing explainable features, called EFC (Explainable Feature Construction), which was designed to work seamlessly for both regression and classification problems. The method involves four steps: 1) explaining of model predictions for individual instances; 2) identifying of groups of attributes that commonly appear together in explanations; 3) efficiently creating of constructs from the identified groups; and 4) evaluating the constructs while selecting the best as new features. The EFC method reduces the time needed to construct features and improves the classification accuracy of several classifiers on standard datasets. Further, in the presented study a domain expert validated the plausibility of the generated features on the real domain. Understanding machine learning models is key to improving their usability and ensuring reliable results. This can be achieved by developing interpretable models and using explanation methods. One of the effective explanation methods are perturbation methods, which provide understanding of a model based on changes in model output that occur when the input data changes. In the dissertation, the use of perturbation methods is presented as a tool to reduce the search space and thereby detect informative groups of attributes. The main contribution of the dissertation is the innovative heuristic method applied for reducing the search space in constructive induction based on explanations of predictive models. The method allows the generation of informative groups of candidate constructs that include groups of attributes that commonly appear in local explanations of black-box models. These groups are used to generate meaningful explainable features that improve the predictive performance of several classifiers. While this method significantly reduces the computational complexity of the search, it also enables the effective involvement of a domain expert, which adds to the comprehensibility of predictive models.20242024-03-20 16:10:07Doktorsko delo/naloga155117sl