Evaluating the performance of bagging-based k-nearest neighbor ensemble with the voting rule selection method

Label ranking based prediction problems help to find a mapping between instances and ranked labels. To improve the prediction performance of label ranking models, the main suggestion is to use ensembles. The main feature of ensemble models is the ability to retrieve the outcomes of various multiple simple models and join them into a combined outcome. To ensure the property of ensemble learning, the nearest neighbor estimation is applied with a bagging approach that samples the instances of training data. The reason to select the label ranking ensemble approach is shown in this study using detailed analysis of suitable existing algorithms. The results show that the defined parameters used in the k-nearest label ranker help to obtain a better prediction performance than the existing label ranking ensembles with an accuracy of 85% to 99% on 21 label ranking datasets. But there is a possibility to improve any ensemble algorithm by using the voting rule selection procedure. This study, integrating the Voting Rule Selector (VRS) algorithm and the seven commonly used voting rules with the k-nearest neighbor label ranker, found that VRS and Copeland work more efficiently than the Borda aggregation in dataset level learning. The results indicate that the k-nearest neighbor label ranker with the VRS and Copeland aggregation method is ranked first in most of the datasets. In the dataset level, the VRS method obtained an average improvement of 48.02% in comparison to the simple model k-nearest neighbor approach and is almost equal to the Copeland with 47.84%.