Effective construction of classifiers with the k-NN method supported by a concept ontology

Abstract

In analysing sensor data, it usually proves beneficial to use domain knowledge in the classification process in order to narrow down the search space of relevant features. However, it is often not effective when decision trees or the k-NN method is used. Therefore, the authors herein propose to build an appropriate concept ontology based on expert knowledge. The use of an ontology-based metric enables mutual similarity to be determined between objects covered by respective concept ontology, taking into consideration interrelations of features at various levels of abstraction. Using a set of medical data collected with the Holter method, it is shown that predicting coronary disease with the use of the approach proposed is much more accurate than in the case of not only the k-NN method using classical metrics, but also most other known classifiers. It is also proved in this paper that the expert determination of appropriate structure of ontology is of key importance, while subsequent selection of appropriate weights can be automated. Keywords k-nearest neighbour algorithm • Ontology similarity metrics • Holter measurement • Coronary disease B Jan Bazan

Assessment measures of an ensemble classifier based on the distributivity equation to predict the presence of severe coronary artery disease

“…To some extent, it also works as a screening tool for all patients with CAD. Similar considerations for other classification methods were conducted by Bazan et al (2020).…”