Research On Random Hierarchical Neural Network Based One-Class Classification

One-class classification(OCC)learning aims to construct a feature description and recognition model from the only target data to realize the effective distinction between the target class and the abnormal class.In the real world,OCC learning has a wide range of application scenarios,such as anomaly detection,fraud detection and network attack detection.As a popular network structure with fast training speed and good generalization performance,stochastic neural network has been widely studied in OCC learning.In particular,in response to the demands of large-scale data,OCC algorithm based on deep stochastic neural network has received a lot of attention.However,most of the current mainstream deep stochastic neural network based OCC are transplanted from multi-classification problems,and they have not specifically optimized or designed directly for OCC problems.As a result,the data representation ability and generalization performance of the network have not been fully developed.Based on this,this paper focuses on the OCC problem,and studies from the aspects of improving the data representation ability,generalization performance and robustness to noise of the hierarchical stochastic neural network.The main research contents are as follows:1.Aiming at the weakness of deep stochastic neural network in feature representation and robustness under OCC learning target,a novel maximum correntropy criterion based hierarchical one-class classification algorithm is proposed.MC-HOC employs stacked WSC-RAE for feature learning and encoding,and adopts MCC to replace MSE as the optimization criterion.WSC-RAE is effective in discriminative feature extraction through imposing constraint on the within-class scatter of target.Then the proposed MC-HOC is further extended to kernel learning to enhance its generalization performance.Experiments on many benchmark and the urban acoustic signal classification datasets are conducted to show the effectiveness of MC-HOC/MC-HKOC.2.Aiming at the shortcomings of WSC-RAE’s insensitivity to within-class scatter and poor interaction ability between samples,and a within-class scatter information based interactive reconstruction AE(WSI-IAE)is proposed.The constraint of WSCRAE on the within-class scatter of encoded features only forces it to be smaller than the original features,ignoring the within-class scatter of encoded features.In addition,WSC-RAE only focuses on self-reconstruction,but samples of the same class have the same characteristics,which is the goal of OCC feature learning.Therefore,WSI-IAE uses mutual reconstruction instead of traditional self-reconstruction,which can learn more essential feature representation and tighter feature distribution from samples.Thanks to this,by stacking multiple WSI-IAEs,it is helpful to obtain the hypersphere with smaller radius,thus greatly improving the performance of deep stochastic neural network based OCC.Finally,the performance of WSI-IAE is demonstrated by comparing several benchmark datasets with various autoencoders.3.An adaptive weighted random autoencoder(AW-RAE)is proposed to solve the problem that WSC-RAE ignores samples in marginal distribution,which resulted in the degradation of representation ability.At present,the algorithm is sensitive to large outliers,since every sample is treated equally during reconstruction,but samples near the edge may cause description deviation.Therefore,in order to weaken the side effects of the samples near the edge,AW-RAE uses a weighted strategy that is only relevant to the nearest neighbors’ distribution knowledge to assign different reconstruction penalties to the training data.Specifically,an adaptive weighted method based on cosine sum and relative density is used to assign lower reconstruction penalty to samples at the edge of the training data.Comparative experiments on benchmark and the urban acoustic signal classification datasets are conducted to show the performance of AW-RAE.