Research On New Hierarchical Fuzzy Classification Learning Method

In recent years,artificial intelligence has been widely used in various fields of human life.In particular,machine learning has made breakthroughs in many knowledge tasks such as classification,clustering and regression.Among them,classification is one of the most important research branches in machine learning.It is widely used in practical applications such as semantic analysis,image recognition,text classification and assistant medical diagnosis.As an important branch of artificial intelligence,Takagi-Sugeno-Kang(TSK)has been applied in many fields.How to improve its classification performance and interpretability is still a challenging task.However,in the face of emerging application scenarios,the following challenges are usually faced:(1)how to extract randomly some features of the original training samples,and how to maintain the classification performance and strong interpretability of the classifier,while also applying to the classification for large scale data;(2)how to improve the generalization performance of each base training block while ensuring the classification performance of the proposed classifier;(3)how to enhance the ability of avoiding over-fitting phenomenon of the classifier;(4)how to solve the shortcomings of the existing hierarchical fuzzy classifiers in interpreting the outputs of the middle layers and fuzzy rules.In order to solve the problems faced by the traditional classifier in dealing with the above application scenarios,this paper mainly improves the classical fuzzy classifier from the perspective of classification to achieve satisfactory classification performance.The specific work is as follows:(1)The first part is the second chapter,which mainly discusses the hierarchical fuzzy classification problem of training only local feature information.In order to further enhance the classification performance and interpretability of the hierarchical fuzzy classifier,this section proposes a hierarchical TSK fuzzy classifier based on the stacked structure principle.The proposed hierarchical fuzzy classifier is constructed block by block using a stacked structure.Each base training block is composed of a zero-order TSK fuzzy classifier.According to the stacked structure principle,the training dataset obtained by the random projection of the prediction result of the current base training block is randomly changed and used as the input of the next base training block.For the proposed hierarchical fuzzy classifier,the hidden layers in each base training block are represented by fuzzy rules with interpretability.These fuzzy rules are selected by random combination and random rule combination.Based on the fixed five fuzzy partitions,the source training dataset is mapped to each of the independent base training blocks as the same input space.The output layer of each base training block can be quickly learned by a least learning machine(LLM).Therefore,the proposed hierarchical fuzzy classifier can be well applied to the classification for large-scale data.Finally,experimental results show that the proposed hierarchical TSK fuzzy classifier has good classification performance and interpretability.(2)The second part is the third chapter,which mainly discusses the hierarchical fuzzy classification problem based on the optimization training features.Aiming at the shortcomings of the existing hierarchical fuzzy classifier in interpreting the outputs of middle layers and fuzzy rules,this section proposes a hierarchical TSK fuzzy classifier with good classification performance and high interpretability.In order to enhance the classification performance,each base training block in the proposed hierarchical TSK fuzzy classifier is organized by a special superposition method.This proposed training method allows all input features of the original training sample plus the output of the previous training block to constitute an optimized training space,which in turn serves as an input to the next training block.These optimization features essentially open the manifold structure of the original input space,enabling enhanced classification performance.When designing each base training block,LLM is used to quickly obtain the analytical solution of the fuzzy rule,which improves the training efficiency.Each fuzzy rule is generated by randomly selecting an input feature and randomly selecting a fixed Gaussian membership function corresponding to the selected input features.The fuzzy rules generated by this method are interpretable.The experimental results also show that the proposed hierarchical TSK fuzzy classifier has enhanced classification performance and high interpretability.(3)The third part is the fourth chapter,which mainly discusses the hierarchical fuzzy classification problem based on the integrated combination in the training block.This section proposes an interpretable TSK fuzzy classifier block combination,which achieves good classification performance and strong interpretability.As a special hierarchical fuzzy classifier,the classifier is built by block-by-block superposition.Each base building block is composed of multiple zero-order TSK fuzzy classifiers.These zero-order TSK fuzzy classifiers also use the method of negative correlation learning(NCL)for block training,which improves the generalization ability of the base training blocks.Using the stacked generalization principle,the output of the current base training block is randomly projected to the next base training block,and then combined with the current training samples.The purpose of constructing this particular hierarchy is to ensure that all base training blocks can be trained in the same input/output space,the current training samples,and the random projection output of the previous training block.In the input layer,the target output of the current training sample is used instead of the random projection output of the previous training block.Each TSK fuzzy classifier in the base training block is composed of interpretable fuzzy rules,which in turn assign a fuzzy subset in the fixed fuzzy partition to each selected input feature by randomly selecting the input features.The final experimental results show that the hierarchical TSK fuzzy classifier does have obvious classification advantages.(4)The fourth part is the fifth chapter,which mainly discusses the hierarchical fuzzy classification problem based on the restricted Boltzmann machine(RBM)and avoids over-fitting phenomenon.This section proposes a new hierarchical TSK fuzzy classifier with good classification performance,high interpretability and avoidance of over-fitting.Different from the previous deep TSK fuzzy classifier,the proposed hierarchical TSK fuzzy classifier takes several zero-order TSK fuzzy classifiers as its base training blocks,and then organizes them into a hierarchical structure to make the current base training blocks.The output plus the corresponding output of the previous base training block is as close as possible to their expected output.In order to ensure the interpretability and fast learning of the proposed classifier,this section projects the original input space onto a randomly selected subset of the original feature set.Each base training block has its own input space and short fuzzy rules.The antecedent is generated by randomly selecting features.This section selects five Gaussian membership functions with a concise language explanation,and uses LLM to quickly learn its corresponding consequent parameters.As an alternative to layer-to-layer overlay,the proposed hierarchical TSK fuzzy classifier is primarily designed from a biological perspective.Because it has a reliable theoretical basis,that is,the training of each base training block can be equivalently understood from the generalized restricted Boltzmann machine(RBM)in the hierarchical TSK fuzzy classifier.This training model proves the superiority of deep RBM in both theory and experience.This section demonstrates for the first time the decisive role played by the proposed classifier in training such hierarchical fuzzy classifiers.Theoretical analysis shows that the constructed hierarchical TSK fuzzy classifier has obvious ability to avoid over-fitting.Finally,the experimental results also show the effectiveness of the hierarchical TKS fuzzy classifier in classification performance,interpretability and avoiding over-fitting.This also means that the TSK fuzzy classifier proposed in this section provides a more flexible and advantageous choice for designing deep TSK fuzzy classifiers.