Temporal Fusion Decoder

Temporal Fusion Decoder
ls-type:: annotation
hl-page:: 10
hl-color:: yellow

Locality Enhancement with Sequence-to-Sequence Layer
ls-type:: annotation
hl-page:: 10
hl-color:: yellow
在 TFD 前用 [[Seq2Seq]] 对数据进行一次增强 #card

• 重要的点通常是与周围的值相关系来确定的，例如异常值、变化点或周期模式 points of significance are often identified in relation to their surrounding values – such as anomalies, change-points or cyclical patterns.
  ls-type:: annotation
  hl-page:: 10
  hl-color:: yellow

• 注意力机制架构中构建利用 point-wise values 的特征可以实现性能提升

  • [[@Enhancing the Locality and Breaking the Memory Bottleneck of Transformer on Time Series Forecasting]] single convolutional layer for locality enhancement

ls-type:: annotation
hl-page:: 10
hl-color:: green

• LSTM encoder-decoder

  • encoder 输入 $\tilde{\boldsymbol{\xi}}_{t-k: t}$ 前 k 步特征

  • decoder 输入 $\tilde{\boldsymbol{\xi}}{t+1: t+\tau{\max }}$ 未来 tao 步特征

  • 由 encoder 和 decoder 生成的时序特征可表示为 $\phi(t, n) \in\left{\boldsymbol{\phi}(t,-k), \ldots, \boldsymbol{\phi}\left(t, \tau_{\max }\right)\right}$

    • 其中 n 为位置索引

  • 时序特征输入到 TFD 前的在经过一次非线性变换

    • $\tilde{\boldsymbol{\phi}}(t, n)=\operatorname{LayerNorm}\left(\tilde{\boldsymbol{\xi}}{t+n}+\operatorname{GLU}{\tilde{\phi}}(\boldsymbol{\phi}(t, n))\right)$

Static Enrichment Layer
ls-type:: annotation
hl-page:: 10
hl-color:: yellow
用静态变量增强时序特征 #card

• 静态协变量通常对时间动态有重要影响 static covariates often have a significant influence on the temporal dynamics
  ls-type:: annotation
  hl-page:: 10
  hl-color:: yellow

• $\boldsymbol{\theta}(t, n)=\operatorname{GRN}_\theta\left(\tilde{\boldsymbol{\phi}}(t, n), \boldsymbol{c}_e\right)$

  • ce is a context vector from a static covariate encoder

ls-type:: annotation
hl-page:: 10
hl-color:: yellow

自关注模块 Temporal Self-Attention Layer
ls-type:: annotation
hl-page:: 11
hl-color:: yellow
学习时序数据的长期依赖关系并提供模型可解释性 #card

• [[Interpretable Multi-Head Attention]]，每个 head 使用相同的 QW，然后用多头 attention score 加权平均后的 V 做解释

  • $\boldsymbol{B}(t)=$ InterpretableMultiHead $(\boldsymbol{\Theta}(t), \boldsymbol{\Theta}(t), \boldsymbol{\Theta}(t))$

• $\boldsymbol{\delta}(t, n)=\operatorname{LayerNorm}\left(\boldsymbol{\theta}(t, n)+\operatorname{GLU}_\delta(\boldsymbol{\beta}(t, n))\right)$.

Position-wise Feed-forward Layer
ls-type:: annotation
hl-page:: 11
hl-color:: yellow
对自关注层的输出应用额外的非线性处理公式 #card

• $\boldsymbol{\psi}(t, n)=\operatorname{GRN}_\psi(\boldsymbol{\delta}(t, n))$

• $\tilde{\boldsymbol{\psi}}(t, n)=\operatorname{LayerNorm}\left(\tilde{\boldsymbol{\phi}}(t, n)+\operatorname{GLU}_{\tilde{\psi}}(\boldsymbol{\psi}(t, n))\right)$