The crossing phenomenon and power of the 1-NN rule

0480210 - ÚI 2025 DE eng J - Článek v odborném periodiku
Jiřina, Marcel - Krayem, S.
The crossing phenomenon and power of the 1-NN rule.
Advances in Data Analysis and Classification. -, accepted after minor revisions 2020 (2024). ISSN 1862-5347. E-ISSN 1862-5355
Grant CEP: GA MŠMT(CZ) LG15047
Institucionální podpora: RVO:67985807
Klíčová slova: kNN rule * multivariate data * classification * distance * nearest neighbor * distribution mapping function
Impakt faktor: 1.6, rok: 2022

The paper brings a detailed insight into one of causes of misclassification in popular k-NN classifiers. As a tool we use the distribution mapping function. This function can be constructed as dependence of the neighbor rank on the distance of the neighbor from the query point. In classification problems if there are two classes, the DMFs for a given query point can be constructed separately for one and the other class. They often lie one above another and thus decision on resulting class of the query point is definite. It is given by the class for which the DMF lies more to the left than the other. However, these functions may cross. We show that this crossing may influence decision of the k-NN rules especially if a relatively large k is used. Our results show that the crossing phenomenon is a source of errors and in 36% of cases its elimination improves behavior of the k-NN rules.
Trvalý link: http://hdl.handle.net/11104/0276079