Attribute Correlation Modeling And Missing Value Imputation Of Incomplete Data Based On Fuzzy Partition

Discovering the laws of things and mining valuable knowledge from mass data has always been the focus of researchers in the field of data mining.But the existence of missing data increases the difficulty of data mining,which will reduce the reliability of analysis results.Filling the missing values reasonably has become a very important link in data analysis and mining.In this paper,the data modeling method is used to fill the missing values in incomplete data,and the correlation of data attributes is mined by constructing the model.With the goal of improving the model's ability to approximate incomplete data attribute correlation,the main studies of this paper are as follows:(1)For the phenomenon that the differences exist in the attribute correlations among different sample categories,the linear regression models of incomplete data attributes are established by the method of class-based modeling.The incomplete data set is divided by the fuzzy C-means clustering algorithm based on the optimized complete strategy(OCS-FCM).And then linear regression models are established for the corresponding attribute relationship on each fuzzy subset.Then,the modeling of incomplete data and imputation of missing value based on Takagi-Sugeno(T-S)structure are realized.(2)For the complexity of the correlation between attributes of incomplete data,a method of incomplete data modeling based on the single output neural network is proposed.The neural network models are constructed in turn by taking the attribute with missing values as input and other attributes as output.This single output neural network structure can accurately depict the relationship between each attribute and other attributes.For the problem of incomplete model input due to existence of missing values,missing values are described as system-level variables.The network parameters and missing values variables are updated alternately through iterative learning among networks.This method can effectively use the information of all observed data in incomplete data to mine the attribute relationship of incomplete data.(3)On the basis of the above two studies,we proposes a modeling method of incomplete data that builds the single output subnet group on T-S fuzzy system.After fuzzy division by OCS-FCM,single output neural network model groups are established for incomplete data on each fuzzy subset,and the corresponding subnet models on each subset are connected smoothly by membership information.Furthermore,the establishment of the TS-subnet Group model of incomplete data and the missing value filling based on the model are realized.The TS-subnet Group modeling method proposed in this paper fully combines the precision of TS modeling and the accuracy of attribute relation fitting with single output subnets.The effectiveness of this method is verified by experimental results.