Back to results
US 11,537,874 B1Granted
Deep factor models with random effects for forecasting
AssigneeAmazon Technologies, Inc.
InventorsYuyang Wang, Alexander Johannes Smola, Dean P. Foster, Tim Januschowski
Life of this patent
Estimating the term…
Abstract
Techniques for forecasting using deep factor models with random effects are described. A forecasting framework combines the strengths of both classical and neural forecasting methods in a global-local framework for forecasting multiple time series. A global model captures the common latent patterns shared by all time series, while a local model explains the variations at the individual level.
Claims20
A computer-implemented method comprising:
- receiving, at a web service endpoint of a provider network, a request to generate a time series forecast for a first entity for a time window comprising one or more points of time;
- generating, using at least a recurrent neural network (RNN) trained based on time series data from a plurality of entities, a global prediction for the time window;
- generating, using at least a state-space model (SSM) corresponding to the first entity, a local prediction for the time window, wherein the SSM was trained to model a difference between one or more global patterns generated by the RNN and values from time series data of the first entity;
- generating, based on the global prediction and the local prediction, the time series forecast for the first entity;
- and
- sending data comprising the time series forecast.
3 more elements
A computer-implemented method comprising:
- generating, using at least a neural network, a global prediction for one or more points in time, wherein the neural network was trained based on time series data of a plurality of entities;
- generating, using at least a probabilistic model that is specific to a first entity, a local prediction for the one or more points in time, wherein the probabilistic model was trained to model a difference between one or more global patterns generated by the neural network and values from time series data of the first entity;
- generating, based on the global prediction and the local prediction, a time series forecast for the first entity;
- and
- sending data comprising the time series forecast.
2 more elements
A system comprising:
- a first one or more electronic devices of a provider network, the first one or more electronic devices comprising one or more processors and instructions that upon execution cause the provider network to:
- receive a request to generate a time series forecast for a first entity for a time window comprising one or more points of time;
- and
- send the request to a forecasting service of the provider network;
- and
- a second one or more electronic devices to implement the forecasting service of the provider network, the forecasting service including instructions that upon execution cause the forecasting service to:
- train a probabilistic model that is specific to the first entity to model a difference between one or more global patterns generated by a neural network and values from time series data of the first entity;
- generate, using at least the neural network, a global prediction for one or more points in time,
- wherein the neural network is trained based on time series data of a plurality of entities;
- generate, using at least the probabilistic model that is specific to the first entity, a local prediction for the one or more points in time, wherein the probabilistic model is trained to model a difference between one or more global patterns generated by the neural network and values from time series data of the first entity;
- generate, based on the global prediction and the local prediction, the time series forecast for the first entity;
- and
- send data comprising the time series forecast.
10 more elements
Dependencies are inferred from the claim text.