Font Size: a A A

Research On TSK Fuzzy Classifier Ensemble Method For Complex Scenes

Posted on:2022-11-20Degree:MasterType:Thesis
Country:ChinaCandidate:X H LiFull Text:PDF
GTID:2518306764999689Subject:Computer technology
Abstract/Summary:PDF Full Text Request
In recent years,with the rapid development of science and technology,various fields in society are actively integrated with artificial intelligence across fields,such as smart campus,intelligent medical system,electronic police and so on.In the context of the rapid development of artificial intelligence,machine learning as the most intelligent in many fields of artificial intelligence,has been successfully and widely applied in many application scenarios,such as recommendation system,facial recognition and disease detection.For traditional machine learning algorithms,application in various application scenarios is both an opportunity and a huge challenge.The different scenarios mentioned in this article refer to data scenarios with different requirements.In supervised machine learning,TSK fuzzy system and traditional neural network often have the following problems when it deals with the data from different application scenarios: When using imbalanced raw data for learning modeling,minority sample is always being ignored.When using complex raw data for learning modeling,the knowledge obtained from complex data is difficult to be explained by real semantics.When using large-scale raw data for learning modeling,the model performance is usually poor and the training time is always too long.To solve these problems,this dissertation will build different integration architectures and improve feature selection algorithms on the basis of the existing classical machine learning field,in order to improve interpretability of TSK fuzzy system and traditional neural network to complex data.The main contents are as follows:(1)New classifier based on Particle Swarm Optimization Feature Selection(PSOFS)and Takagi-Sugeno-Kang(TSK)fuzzy system is proposed,i.e.,parallel ensemble fuzzy neural network based on PSOFS and TSK(PE?PT?FN),is used for imbalance data.Each class samples in the training set are randomly sampled,and those samples obtained by being randomly sampled are added.Then,the feature selection method PSOFS is carried out independently and parallelly.In PSOFS,particles which are random initial positions represent different feature subsets and converge to the optimal positions after many iterations.Each subset has a corresponding feature subset.Several groups of TSK fuzzy neural networks(TSK?FN)are trained by each feature subset in parallel.In imbalanced data scenarios,PE?PT?FN can efficiently acquire knowledge from data.(2)A novel deep ensemble TSK fuzzy classifier based on feature selection of multi-label and mutual information(FMMI?TSK)is proposed.Firstly,feature selection of multi-label and mutual information(FMMI)was used to obtain the set of feature subsets containing different feature information from data.In this set,the relevance between feature subset and the output results of corresponding label is maximum,and the redundancy between feature subsets is minimum.Secondly,the labels of each feature subset are reset.Since the feature information in each feature subset corresponds to a certain type of label,the labels of the data samples belonging to this type of label are set to 1,and the labels of the data samples belonging to other labels are set to 0.Redundant data are randomly deleted to obtain the feature subset of balance data.Then,the classical neural network algorithm Radial Basis Function neural network is used to obtain the membership function from the feature subset,so as to obtain the premise parameter of fuzzy rules.At last,the algorithm of Extreme Learning Machine is used to obtain the consequence parameter from the fuzzy membership value.FMMI?TSK can extract knowledge from complex data in a way that humans could understand.(3)Parallel ensemble fuzzy neural network based on K-d Tree and KNN(PE?TSK?FN)is proposed.Firstly,the K-d Tree algorithm is used to divide the original data space into multiple independent rectangular subspaces at the level of data space.In other words,the original data is divided into multiple subsets.Then,these data sets are used to train independent fuzzy sub-classifiers in parallel.Finally,K nearest neighbor sub-classifiers selected from the sub-classifiers through KNN algorithm to vote and the final output is obtained.In large-scale data scenarios,the ensemble structure of PE?TSK?FN can not only accelerate the training speed of the model,but also avoid the over-fitting phenomenon of the model.
Keywords/Search Tags:Ensemble classifier, TSK fuzzy classifier, Imbalance data, Interpretability, Large-scale data
PDF Full Text Request
Related items