Amazon Technologies, Inc.

Jan Alexander Gasthaus

Tim Januschowski

Valentin Flunkert

Matthias Seeger

Lorenzo Stella

Sebastian Schelter

Yuyang Wang

Gregory Michael Duncan

Konstantinos Benidis

David Salinas

Christian Uriel Carmona Perez

Syama Rangapuram

Joos-Hendrik Boese

Francois-Xavier Benoit Marie Aubet

Mohamed El Fadhel Ayed

Techniques for anomaly detection are described. An exemplary method includes receiving a request to monitor for anomalies from one or more data sources; analyzing time-series data from the one or more data sources; generating a recommendation for handling the determined anomaly, the recommendation generated by performing one or more of a root cause analysis, a heuristic analysis, and an incident similarity analysis; and reporting the anomaly and recommendation.

Systems, apparatuses, and methods for anomaly detection

Systems and methods are described for detecting anomalies within data, such as time series data. In one example, unlabeled data, such as time series data, may be obtained. At least one data point, representing an artificial anomaly, may be inserted into the data. The data may then be divided into a number of different windows. The windows may have a fixed size and may at least partially overlap in time. The data contained within different windows may be compared, to each other and to the injected data point, to determine an anomaly score for individual windows. The anomaly score may indicate a likelihood that a given window contains an anomaly. In a specific example, a convolution neural network may be trained based on the data and inserted data points representing anomalies, where a contrastive loss function is used to represent different portions of the data in the neural network.

Neural contrastive anomaly detection

Techniques are described for a time series probabilistic forecasting framework that combines recurrent neural networks (RNNs) with a flexible, nonparametric representation of the output distribution. The representation is based on the nonparametric quantile function (instead of, for example, a parametric density function) and is trained by minimizing a continuous ranked probability score (CRPS) derived from the quantile function. Unlike methods based on parametric probability density functions and maximum likelihood estimation, the techniques described herein can flexibly adapt to different output distributions without manual intervention. Furthermore, the nonparametric nature of the quantile function provides a significant boost in the approach's robustness, making it more readily applicable to a wide variety of time series datasets.

Probabilistic forecasting with nonparametric quantile functions

A composite time series forecasting model comprising a neural network sub-model and one or more state space sub-models corresponding to individual time series is trained. During training, output of the neural network sub-model is used to determine parameters of the state space sub-models, and a loss function is computed using the values of the time series and probabilistic values generated as output by the state space sub-models. A trained version of the composite model is stored.

Artificial intelligence system combining state space models and neural networks for time series forecasting

An input data set with a plurality of item descriptors comprising respective time series observations is identified. A routing directive indicating a predicate to be evaluated to determine whether a particular item descriptor is to be included in a training data set for a first learning algorithm is obtained. A plurality of learning algorithms are trained using training data sets derived from the input data set according to respective routing directives, and the trained algorithms are stored.

Training data routing and prediction ensembling at time series prediction system

With respect to an input data set which contains observation records of a time series, a statistical model which utilizes a likelihood function comprising a latent function is generated. The latent function comprises a combination of a deterministic component and a random process. Parameters of the model are fitted using approximate Bayesian inference, and the model is used to generate probabilistic forecasts corresponding to the input data set.

Growing community of inventors

Berlin, Germany

Jan Alexander Gasthaus