Research On Sparse And Heterogeneous Fuzzy Systemm Odeling And Clustering Methods

Takagi-Sugeno-Kang(TSK)fuzzy systems usually adopts a large number of rules for highly nonlinear modeling tasks,which further weaken the clarity and interpretability of TSK FS.So this paper studies how to model with sparse fuzzy learning.With the increasing amount of heterogeneous data,it is necessary to study efficient heterogeneous fuzzy systems.Among them,the multi-view heterogeneous data is a typical class of heterogeneous data.For this purpose,this topic has made a preliminary discussion and explored multi-view clustering techniques suitable for heterogeneous multi-view fuzzy modeling.Multi-view data belongs to different dimension spaces.How to fully mine the relationship between data to improve the clustering performance based on the characteristics of multi-view data is a key challenge for multi-view heterogeneous clustering.For the problems of sparse fuzzy modeling,this paper proposes Concise Monotonic TSK Fuzzy System for monotonic classification(CM-TSK-FS)and Enhanced soft subspace clustering and sparse learning based Concise TSK fuzzy system(ESSC-SL-CTSK-FS).First,for CM-TSK-FS,the algorithm improves the interpretability of the fuzzy system for monotonic data sets.The proposed method has the following advantages: 1)The new method reduces the complexity of TSK fuzzy system due to the feature selection of monotonic data which makes the fuzzy system more concise.2)Since the monotonicity between the features and the decision values of the monotonic data is taken into account when extracting the features,the classification performance of the training model has been improved to a certain extent.Secondly,For ESSC-SL-TSK-FS,the antecedents are generated using ESSC,which can generate the sparse and different fuzzy subspace for different fuzzy rules;Furthermore,the sparse learning technique is used to sparse the consequent parameters of fuzzy rules and finally the number of fuzzy rules can be effectively reduced based on the sparse consequent parameters.Therefore,with ESSC-SL-CTSK-FS,the concise 0-order TSK FS can be constructed,which makes the TSK FS more clear,transparent and easy to explain in the scene of high dimensional data.And this method is extended to a neural network to get a sparse neural network.the Least Absolute Shrinkage and Selection Operator(LASSO)strategy is used to select the hidden layer nodes,and the shrinkage parameter are determined by cross validation and grid search.Therefore,RBF neural network based on Lasso sparse learning(RBF-NN-LASSO)is proposed.The experimental results show that RBF neural network modeling based on Lasso sparse learning can not only reduce the complexity of the model,but also improve the classification accuracy.For the modeling of heterogeneous fuzzy systems,a preliminary discussion was made.This paper mainly discusses a multi-view heterogeneous clustering algorithm that suitable for multi-view heterogeneous fuzzy modeling,namely a hidden space sharing multi-view fuzzy clustering(HHS-MVFC).This method is based on the FCM clustering framework.It introduces hidden space learning mechanism based on multi-view sharing to obtain important correlation information between views.