Research On Anomaly Detection Model Based On Deep Variational Learning

Anomaly detection is an important task in data mining.Its basic goal is to detect data values that deviate from the overall data characteristics.With the development of information technology,existing anomaly detection models have been unable to meet the current anomaly detection needs.Based on deep learning Anomaly detection model has gradually become a new research direction.There are two main types of models in the field of deep anomaly detection.One is the hybrid model.The hybrid model is a model that combines traditional anomaly detection algorithms and deep learning dimensionality reduction algorithms to improve detection efficiency.The second is to generate a model,which directly models the data through training data,estimates its distribution characteristics by observing samples,and combines the discriminant model to detect anomalies in high-dimensional data.Aiming at the advantages and disadvantages of the current anomaly detection models,a hybrid detection model is proposed.The anomaly detection is performed by combining deep learning dimensionality reduction algorithms and improved traditional detection models.The main work of this paper is as follows:(1).Discussion and analysis of anomaly detection related workIt summarizes the current development of deep learning in the field of anomaly detection,and summarizes the current status of related technologies and data reduction in the field of anomaly detection.A hybrid model based on deep variational dimensionality reduction is proposed and used for anomaly detection,and its advantages and disadvantages are outlined,and the research direction is clarified.(2).Local anomaly factor algorithm based on two-way neighbor correctionThis paper proposes a local outlier factor algorithm based on bidirectional neighbor correction.The bidirectional neighbor search algorithm is used to select better parameters for calculating outliers.The pruning algorithm has been proposed to reduces the neighborhood searching time and unnecessary outlier's calculations.It also comprehensively uses the latest Neighbors and reverse nearest neighbors introduce a bidirectional neighbor-based correction factor and use reverse neighbors to further improve detection accuracy.Finally,in the experimental evaluation of synthetic data set and UCI data set,the algorithm has better performance in parameter selection and time efficiency than other models.(3).Anomaly detection model based on deep variation dimensionality reductionFirst,the advantages and disadvantages of auto-encoders and other related generative models are outlined,and the application of variational auto-encoders in the field of data dimension reduction is described.Secondly,based on the idea of full dimensionality reduction and variational autoencoder,a deep variational dimensionality reduction model is proposed,which is combined with a two-way neighborhood algorithm to form a hybrid model.Finally,comparative performance tests were performed on MNIST and CIFAR-10 data sets,and the results show that the proposed model has advantages over other hybrid and generative models.Finally,the paper summarizes and prospects.