【论文阅读】A Dual-Stage Attention-Based Recurrent Neural Network

摘要

Nonlinear autoregressive exogenous (NARX) model 用于 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

然而目前的研究问题主要在于：① 捕捉长时间的时序依赖性问题 ② 选择决策的relevant driving series

提出了模型a dual-stage attention-based recurrentneuralnetwork (DA-RNN)：

① input attention mechanism adaptively extract relevant driving series at each time step

② temporal attention mechanism to select relevant encoder hidden states across all timesteps

Introduction

时间序列分析模型沿革

well-known autoregressive moving average (ARMA) model 及其变体—— 不能模拟非线性的关系以及区分exogenous (driving) input terms

kernel methods、ensemble methods、Gaussian processes ——使用了预先设置的nonlinear form，