Performance Analysis of Class Imbalance in Link Quality Estimation
Mu He, Jian Shu, Linlan Liu, Mingxiao Niu · 2021
In the wireless sensor networks, one of the common reasons that affect the accuracy of link quality estimation models is class imbalance. Machine learning-based methods focus on maximizing the total accuracy over the entire dataset leading to more attention being paid to the majority class samples. Hence, the minority class samples are poorly predicted by the learning model. This paper aims to study the effect of class imbalance on link quality estimation models. The stratified sampling algorithm is used to change the class imbalance ratio to analyze the relationship between the estimator and the class imbalance ratio. We carry out the experiments in real campus parking lots scenario and office scenario. The experimental results show that the relationship between the precision of linear estimator and the class imbalance ratio is concave. The precision of nonlinear estimator is best when the link quality sample is imbalanced.