PP: A dual-stage attention-based recurrent neural network for time series prediction

Problem: time series prediction

The nonlinear autoregressive exogenous model: The Nonlinear autoregressive exogenous (NARX) model, which predicts the current value of a time series based upon its previous values as well as the current and past values of multiple driving (exogenous) series.

However, few NARX models can capture the long-term temporal dependencies appropriately and select the relevant driving series to make a prediction.

2 issues:

1. capture the long-term temporal dependencies

2. select the relevant driving series to make a prediction

We propose a dual-stage attention-based RNN to address these 2 issues.

1. first stage: input attention mechanism to extract relevant driving series.

2. second stage: temporal attention mechanism.

attention-based encoder-decoder networks for time series prediction/ LSTM/ GRU

One problem with encoder-decoder networks is that their performance will deteriorate rapidly as the length of input sequence increases.

Contribution: the two-stage attention mechanism. input attention for driving series and temporal attention for all time stamps.

input attention can select the relevant driving series.

temporal attention capture temporal information.

Supplementary knowledge:

1. what is driving series?