Study On Fuzzy Clustering For Stylistic Data And Fuzzy Deep Learning

Unsupervised learning and supervised learning have been the two important research areas in machine learning.What is more,the clustering and classification technologies have been applied to military,industrial and civil applications such as intelligent medical treatment,image recognition,natural language processing,video analysis,industrial detection,intelligent transportation,and so on.Especially in the past decade,the artificial intelligent technologies which are significantly represented by deep learning techniques have developed into a new height.However,with the development of various clustering and classification techniques,how to reasonably learn and utilize the features contained in data samples have always been hot fields.Specially,it is related to the following three aspects:1)most of the existing machine learning models are trained by using the physical features(e.g.,distance,color or similarity)of data.However,data samples coming from a same data class present their distinct topologic structure which differs from those from other data classes.The topologic structures of data samples from different data classes obviously present or potentially own style information of data which are significantly different from the physical features of data.2)disturbance or noise samples actually exist in most of real-world datasets,and without the consideration of the noise samples in training stage,the performance of a machine learning model trained by the training samples containing the noise samples will be greatly affected.3)the“rule explosion”problem will occur when using the traditional Takagi-Sugeno-Kang(TSK)fuzzy classifier to solve the classification problems on large-scale datasets.Meanwhile,the generalization capability and the interpretability of the corresponding TSK fuzzy system will be significantly degraded by the arising“rule explosion”problem.With regard to the above issues,we have carried on some related researches on data modeling for the fuzzy clustering on stylistic data and fuzzy deep learning.The main research results are stated as follows.(1)In view of the fact that data samples originating from different data classes in a real-world dataset obviously present or potentially possess style information of data,a novel classification method Fu CM based on physical and implicit style features of data is proposed.1)In training stage,by the proposed classification method,a social network corresponding to the given dataset is firstly built based on graph theory,in which each training sample and each data class correspond to the node and the subnetwork,respectively.Based on the built social network,a fuzzy social network can be obtained by fuzzifying the weight of each node based on the introducing fuzzy logic technique,and then the style information of data,i.e.,the authority and the fuzzy influence of each node can be naturally exploited.Especially,the fuzzy influence of each node can be iteratively calculated based on its local density and resulting in the corresponding dynamics.2)In prediction stage,a testing sample gets its predicted label by the proposed classification method based on both physical and implicit style features of data.(2)A social-network-based stylistic data classification method called S~2CM is proposed from the perspective of simulating human behavior for perceiving an object.Similarly,in training stage,S~2CM aims at exploring the topologic structure of the social network built according to the given training samples and then exploiting the style information of data in the built social network,namely the authority and the influence of each node.In prediction stage,S~2CM reasonably labels the testing samples by double similarity assumptions,i.e.,the physical similarity assumption and the stylistic similarity assumption.By defining the distance between each pair of nodes in the built social network,the classification behavior of S~2CM based on the characteristic of human cognitive way of an object essentially becomes equivalent to Bayes decision rule.(3)Based on(1)and(2),as the first attempt and research,we aim at designing a fuzzy style clustering method called S-KPC for style data such that samples sharing a same style are grouped into a cluster.S-KPC can distinguish the nuances between different styles of samples originating from different clusters and guarantee the clustering performance from two aspects.Specially,S-KPC firstly expresses the data structures of clusters and mathematically determines the corresponding style information of data by style matrices correspond to the clusters.Then,S-KPC incrementally generates enhanced nodes based on the original features of samples to gradually move the manifold structure of data apart which can make the original samples become easily be clustered.(4)Based on intentionally adversarial samples learning,a novel deep fuzzy classifier termed as DSA-FC with a novel stacked structure is proposed.1)By the conduction of intentionally adversarial sample learning,it means that we will make adversarial samples by intentional attacks on the labels of a small number of samples which are randomly chosen from the given training dataset.Then the Takagi-Sugeno-Kang(TSK)fuzzy classifier is trained based on the training dataset which contains the adversarial samples.2)DSA-FC has its solid theoretical foundation based on the intentionally adversarial samples learning.3)DSA-FC has high interpretability and excellent generalization capability based on the intentionally adversarial samples learning as well as the stacked generalization principle.(5)When the DSA-FC in(4)is applied to solve the classification problems on large-scale datasets,the“rule explosion”easily happens on DSA-FC which will significantly degrade the interpretability of DSA-FC.Meanwhile,the training speed of DSA-FC becomes very slow.With the guarantee of same or comparable generalization capability of DSA-FC in(4),a fast training algorithm of DSA-FC is developed.1)Based on conducting the gradient guided learning,FTA can make an apparent reducing of the total number of fuzzy rules which are used to train the DSA-FC.Then,the training speed of DSA-FC becomes fast and the corresponding interpretability of the deep fuzzy system is naturally enhanced.2)The staked structure of DSA-FC trained by FTA is built based on the simplified fuzzy rules.3)The gradient guided learning not only provides extra generalization capability for DSA-FC but also brings extra benefit in avoiding the co-adaption between each fuzzy rule.