Identifying anomalies can be a goal in itself (anomaly detection) or a means of improving the performance of other time series tasks, such as predictions. Recent deep learning-based anomaly detection and prediction methods typically assume that the proportion of anomalies in the training data is small enough to be ignored and treat unlabeled data as coming from a normal data distribution. We propose a simple and effective technique to augment existing time series models so that they explicitly consider anomalies in the training data. By increasing the training data with a potential anomaly indicator variable whose distribution was inferred when training the base model using Monte Carlo EM, our method inferred outliers while improving the model's performance on nominal data. We demonstrated the effectiveness of the method by combining it with a simple feed-forward predictive model. We investigated how anomalies in the training set affect the training of predictive models, which are commonly used for anomaly detection of time series, and show that our method improves the training of models.
《Monte Carlo EM for Deep Time Series Anomaly Detection》
Address: http://arxiv.org/abs/2112.14436v1
