Algorithm 1 Attention-Weighted Component-to-Latent Mapping

Notation:

C:                Number of input components (e.g., 53)

$D_{\mathrm{enc}}$:            Encoder LSTM hidden size (e.g., 256)

$D_{\mathrm{main}}$:          Main LSTM hidden size (e.g., 200)

K:                Number of windows (e.g., 7)

T:                Number of time steps per window (e.g., 20)

Require:

$W_{\mathrm{enc}}$: Encoder weight matrix $\in {\mathbb{R}}^{D_{\mathrm{enc}}\times C}$

$W_{\mathrm{main}}$: Main LSTM weight matrix $\in {\mathbb{R}}^{D_{\mathrm{main}}\times D_{\mathrm{enc}}}$

${\alpha }_{\mathrm{win}}$: Window-level attention matrix $\in {\mathbb{R}}^{K\times T}$

${\alpha }_{\mathrm{global}}$: Global attention vector $\in {\mathbb{R}}^K$

Ensure:

M: Component-to-latent mapping matrix $\in {\mathbb{R}}^{D_{\mathrm{main}}\times C}$

1:   procedure ComponentToLatentMapping($W_{\mathrm{enc}},W_{\mathrm{main}},{\alpha }_{\mathrm{win}},{\alpha }_{\mathrm{global}}$)

2:        $M\leftarrow \mathbf{0}$                                                                                     ▹ Initialize M to zeros

3:        for $k\leftarrow 1$ to K do                                                         ▹ Iterate over each window

4:               $W_{\mathrm{main}}^{\left(\mathrm{scaled}\right)}\leftarrow {\alpha }_{\mathrm{global}}\left[k\right]·W_{\mathrm{main}}$                                ▹ Scale by global importance

5:               for $t\leftarrow 1$ to T do                          ▹ Iterate over each time step in window k

6:                      $W_{\mathrm{enc}}^{\left(\mathrm{scaled}\right)}\leftarrow {\alpha }_{\mathrm{win}}[k,t]·W_{\mathrm{enc}}$                     ▹ Scale by temporal importance

7:                      $\mathrm{contribution}\leftarrow W_{\mathrm{main}}^{\left(\mathrm{scaled}\right)}\times W_{\mathrm{enc}}^{\left(\mathrm{scaled}\right)}$         ▹ Compute transformation path

8:                      $M\leftarrow M+\mathrm{contribution}$                    ▹ Accumulate weighted contribution

9:             end for

10:        end for

11:        return M

12:   end procedure