Search

Detecting anomalies accurately in time series data has been receiving considerable attention due to its enormous potential for a wide array of applications. Numerous unsupervised anomaly detection methods for time series have been developed because of the difficulty of obtaining accurate labels. However, most existing unsupervised approaches suffer from the problem of anomaly contamination, which results in models that are unable to learn the normal pattern well and further deteriorate the performance of detection methods. To this end, a novel unsupervised method, called Self-adversarial Variational Autoencoder with Spectral Residual (SaVAE-SR), is introduced for time series anomaly detection in this paper. The SaVAE-SR first produces labels for unlabeled training data using the spectral residual technique to identify the most critical anomalies. A VAE model with a modified loss that can leverage label information to remove the influence of anomalous points is then trained in a self-adversarial manner, enabling the model to self-evaluate the learning of complex data distribution and improve itself accordingly. Specifically, the encoder acts as an encoder to approximate the posterior of latent variables and as a discriminator to evaluate the generative ability of the generator and improve itself accordingly. The generator is trained to capture the underlying data distribution and attempts to produce real samples to deceive the discriminator. The encoder and generator of the model compete with each other just like the behavior of GANs but work together under the theoretical framework of VAEs. As a result, the SaVAE-SR model combines the respective strengths of the VAE and adversarial training but does not require an additional discriminator, which makes the whole model very compact. Extensive experiments on five datasets demonstrate the superiority of the proposed method over the existing state-of-the-art methods. [ABSTRACT FROM AUTHOR]

We address the problem of unsupervised anomaly detection for multivariate data. Traditional machine learning based anomaly detection algorithms rely on specific assumptions of normal patterns and fail to model complex feature interactions and relations. Recently, existing deep learning based methods are promising for extracting representations from complex features. These methods train an auxiliary task, e.g., reconstruction and prediction, on normal samples. They further assume that anomalies fail to perform well on the auxiliary task since they are never trained during the model optimization. However, the assumption does not always hold in practice. Deep models may also perform the auxiliary task well on anomalous samples, leading to the failure detection of anomalies. To effectively detect anomalies for multivariate data, this paper introduces a teacher-student distillation based framework Distillated Teacher-Student Network Ensemble (DTSNE). The paradigm of the teacher-student distillation is able to deal with high-dimensional complex features. In addition, an ensemble of student networks provides a better capability to avoid generalizing the auxiliary task performance on anomalous samples. To validate the effectiveness of our model, we conduct extensive experiments on real-world datasets. Experimental results show superior performance of DTSNE over competing methods. Analysis and discussion towards the behavior of our model are also provided in the experiment section. [ABSTRACT FROM AUTHOR]