Application Of Neural Processes In Anomaly Detection Of Multidimensional Data

In the field of data mining,the emergence of massive data has made anomaly detection methods one of the research focuses in the computer field.In the past few decades,many anomaly detection methods have been proposed and successfully applied in various fields.Due to the powerful modeling capabilities of neural networks,many studies have applied neural networks to anomaly detection methods.Despite their good performance,these methods still face two major challenges.Firstly,it is difficult to obtain labeled data in anomaly detection tasks,and insufficient labeled data limit the flexibility of deep models; Secondly,when the test data have a deviation from the training data,the model usually cannot accurately predict,resulting in uncertainty in the detection results.Therefore,the model needs to have the ability to estimate the uncertainty of the detection results.This thesis proposes a new deep learning anomaly detection framework ADNP to solve the above problems.ADNP incorporates neural processes into the paradigm of anomaly detection,enabling models to effectively utilize labeled and unlabeled data for anomaly detection during training.Unlike other anomaly detection methods,the anomaly detection method proposed in this thesis is equivalent to modeling the distribution of functions representing anomaly patterns,rather than just learning a single neural network function for anomaly detection.The model can predict multiple anomaly detection results for test data by sampling different latent vectors,and can estimate the uncertainty of the model’s anomaly detection results from these results.A large number of experiments on real-world datasets show that ADNP can improve the prediction ability of anomaly detection models and effectively measure the uncertainty of detection results.In this thesis,the proposed method is first implemented based on a multi-layer perceptron.Then,in order to alleviate the posterior failure problem in variational auto-encoders,the Gaussian Copula and the vector quantization coding method are introduced on the basis of ADNP to obtain a model GCVQ-ADNP.Experiments on multiple datasets have shown that both of these methods can alleviate the posterior collapse problem which often occurs in variational auto-encoders,improve the representation ability of the model for latent features,and thus improve the anomaly detection ability of the model.