PP: Think globally, act locally: A deep neural network approach to high-dimensional time series forecasting

Problem: high-dimensional time series forecasting

?? what is "high-dimensional" time series forecasting?

one dimension for each individual time-series. n个time series为n维。

A need for exploiting global pattern and coupling them with local calibration校准 for better prediction.

However, most are one-dimensional forecasting.

one-dimensional forecasting VS high-dimensional forecasting:

1. a single dimension forecast mainly depends on past values from the same dimension.

DeepGLO: a deep forecasting model which thinks globally and acts locally.

A hybrid model: a global matrix factorization model regularized by a temporal convolution network + a temporal network that capture local properties of each time-series and associated covariates相关协变量.

Environment: different time series can have vastly different scales without a priori normalization or rescaling.

Introduction:

需求：比如零售商，one may be interested in the future daily demands for all items in a category. This leads to a problem of forecasting n time-series.

Traditional methods: focus on one time-series or a small number of time-series at a time.

AR, ARIMA, exponential smoothing and so on.

?? how to share temporal patterns in the whole data-set while training and prediction?

RNN - sequential modeling; and suffer from the gradient vanishing/ exploding problems.

LSTM 解决了上述问题。

Wavenet model: temporal convolutions/ causal convolutions.

Temporal convolution has been recently used, however, they still have two important shortcomings:

1. hard to train on data-sets that have wide variation in scales.

2. even though these deep models are trained on the entire data-set, during prediction the models only focus on local past data. i.e only the past data of a time-series is used for predicting the future of that time-series.

global properties. take in multiple time-series in the input layer thus capturing global properties.