Defined
 Supervised learning algorithm that forecasts onedimensional (scalar) time series using recurrent neural networks (RNN)
 Trains a single model jointly over all of the similar time series in your dataset
 Use your trained model to create forecasts for new time series that are similar to the ones on which it was trained
 Training input is one or more target time series that have been generated by the same process or similar processes
 Uses an approximation of the process(es) and predicts how the target time series evolves
Use Cases
 Forecasting global food demand
 Predicting global temperature
 Predicting fastfood queues
 Predicting rainfall
 Predicting stock prices
 Options pricing
 Disease progression
 Supply chain management
SageMaker Algorithms
DeepAR
 Uses a training dataset and an optional test dataset
 Use a trained model to predict forecasts for the future of the time series in the training set, or for other time series
 Automatically derives time series based on the frequency of the target series
 For example: a day series generates
dayofweek
,dayofmonth
,dayofyear
 For example: a day series generates
 Can predict “what if ?” scenarios, such as, what if I change/broaden the distribution of my product ?
 Builds a single model for all timeseries and tries to identify similarities across them
 Important Hyperparameters

context_length
: the number of timepoints that the mode gets to see before making the prediction 
epochs
: the maximum number of passes over the training data 
prediction_length
: the number of timesteps that the model is trained to predict, also called the forecast horizon 
time_freq
: the granularity of the time series in the dataset

本文由
Oscaner
创作, 采用
知识共享署名4.0
国际许可协议进行许可
本站文章除注明转载/出处外, 均为本站原创或翻译, 转载前请务必署名

Previous
[MLSC01] [Algorithms] Reinforcement Learning Algorithms 
Next
[MLSC01] [IMPL and OPs] Introduction