A Wide Fuzzy Apriori Classifier with Triple Diversity Guarantee from Feature, Sample and Rule Levels
Shitong Wang, Runshan Xie · 2024
While fuzzy Apriori method (FAM) features its linguistic interpretability and uncertainty-handling ability, it still has difficulties in computational complexity and generalization capability, caused by the curse-of-dimensionality of input data. In order to overcome these two drawbacks, this study proposes a novel wide ensemble (WTD-FAM) of FAM sub-classifier with triple diversity guarantee. WTD-FAM can be realized as follows. Firstly, WTD-FAM draws several data subsets from input data by using both random subspace method and improved bagging. Then, WTD-FAM constructs FAM sub-classifiers on their respective data subsets, and discards some fuzzy rules in the ruleset of each FAM sub-classifier by knowledge oblivion. Finally, WTD-FAM obtains the final outputs by taking majority voting to aggregate the outputs of all the FAM sub-classifiers. WTD-FAM has three advantages: (1) lower computational complexity by building up each FAM sub-classifier on its own data subset; (2) more diversity among FAM sub-classifiers by triple diversity guarantee from feature, sample and rule levels. (3) extra generalization capability by wide ensemble learning method and knowledge oblivion. Both experimental results and ablation study on eight benchmarking datasets verify the effectiveness of WTD-FAM in the sense of classification performance and training speed.