Open AccessLink quality estimation based on over-sampling and weighted random forest
Author(s) -
Linlan Liu,
Yi Feng,
Gao Sheng-rong,
Jian Shu
Publication year - 2022
Publication title -
computer science and information systems
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.244
H-Index - 24
eISSN - 2406-1018
pISSN - 1820-0214
DOI - 10.2298/csis201218041l
Subject(s) - random forest , computer science , link (geometry) , decision tree , variance (accounting) , quality (philosophy) , tree (set theory) , data mining , network packet , estimation , logistic regression , bayesian probability , statistics , artificial intelligence , pattern recognition (psychology) , machine learning , mathematics , computer network , mathematical analysis , philosophy , accounting , management , epistemology , economics , business
Aiming at the imbalance problem of wireless link samples, we propose the link quality estimation method which combines the K-means synthetic minority over-sampling technique (K-means SMOTE) and weighted random forest. The method adopts the mean, variance and asymmetry metrics of the physical layer parameters as the link quality parameters. The link quality is measured by link quality level which is determined by the packet receiving rate. K-means is used to cluster link quality samples. SMOTE is employed to synthesize samples for minority link quality samples, so as to make link quality samples of different link quality levels reach balance. Based on the weighted random forest, the link quality estimation model is constructed. In the link quality estimation model, the decision trees with worse classification performance are assigned smaller weight, and the decision trees with better classification performance are assigned bigger weight. The experimental results show that the proposed link quality estimation method has better performance with samples processed by K-means SMOTE. Furthermore, it has better estimation performance than the ones of Naive Bayesian, Logistic Regression and K-nearest Neighbour estimation methods.