The Design And Research Of Deep Cascade Fuzzy Decision Forest

In recent years,with the great success of deep neural networks,more and more researchers in machine learning field have been attracted,and various work on deep learning or its application has been proposed.Despite the excellent performance of deep neural network in many real problems,there are still several disadvantages such as inability to be applied to small-scale data,high demand for computing hardware in the training process,and massive model parameters that are difficult to train.Therefore,Zhou et al.proposed a deep cascade forest model based on random decision tree,which to some extent well compensated for these defects.However,in the work of deep cascade forest,the data was not considered as fuzzy data.There are a large number of fuzzy phenomena in real life,and most of the data and knowledge are fuzzy and uncertain.Therefore,in this paper we proposed a deep cascade fuzzy forest model.The detailed work are summarized as follows:First,several classical fuzzy decision tree algorithms were introduced and compared in detail.And the classification performance of these algorithms were experimentally compared.Based on the comparison,the FuzzyID3 algorithm with best classification performance was selected as the basic algorithm for constructing fuzzy decision trees.Second,the fuzzy decision trees obtained were integrated into a fuzzy random forest.The detailed algorithm flow and different combination strategies of fuzzy random forest results were given.Third,based on the idea of deep cascade forest proposed by Zhou,we proposed the learning model of deep cascade fuzzy decision forest.In the paper,we showed the whole process of building the cascade model,and analyzed its advantages over deep neural network.At last,the experimental results on data sets including YEAST,CMC and Vehicle were given to verify the performance of our proposed model.Meanwhile,we also showed the difference of classification results caused by different number of layers in the model.Compared with selected several other tree-based classification models,the results showed that our method was feasible and effective.