1
$\begingroup$

I have the following time-series aggregated input for an LSTM-based model:

x(0): {y(0,0): {a(0,0), b(0,0)}, y(0,1): {a(0,1), b(0,1)}, ..., y(0,n): {a(0,n), b(0,n)}}
x(1): {y(1,0): {a(1,0), b(1,0)}, y(1,1): {a(1,1), b(1,1)}, ..., y(1,n): {a(1,n), b(1,n)}}
...
x(m): {y(m,0): {a(m,0), b(m,0)}, y(m,1): {a(m,1), b(m,1)}, ..., y(m,n): {a(m,n), b(m,n)}}

where x(m) is a timestep, a(m,n) and b(m,n) are features aggregated by the non-temporal sequential key y(m,n) which might be 0...1,000.

Example:

0: {90: {4, 4.2}, 91: {6, 0.2}, 92: {1, 0.4}, 93: {12, 11.2}}
1: {103: {1, 0.2}}
2: {100: {3, 0.1}, 101: {0.4, 4}}

Where 90-93, 103, and 100-101 are aggregation keys.

How can I feed this kind of input to LSTM?

Another approach would be to use non-aggregated data. In that case, I'd get the proper input for LSTM. Example:

Aggregated input:

0: {100: {3, 0.1}, 101: {0.4, 4}}

Original input:

0: 100, 1, 0.05
1: 101, 0.2, 2
2: 100, 1, 0
3: 100, 1, 0.05
4: 101, 0.2, 2

But in that case, the aggregation would be lost, and the whole purpose of aggregation is to minimize the number of steps so that I get 500 timesteps instead of e.g. 40,000, which is impossible to feed to LSTM. If you have any ideas I'd appreciate it.

$\endgroup$
1
$\begingroup$

Sounds to me like you can reshape the input of a timestep as matrix s.t. $X_t \in \mathbb{R}^{n \times 2}$ (assuming you only have a and b as features) and use Convolutional LSTM layers:

$t=0$:
$X_0 = \begin{bmatrix}a(0,0)&b(0,0)\\a(0,1)&b(0,1)\\\vdots&\vdots\\a(0,n)&b(0,n)\end{bmatrix}$

$t=m$:
$X_m = \begin{bmatrix}a(m,0)&b(m,0)\\\vdots&\vdots\\a(m,n)&b(m,n)\end{bmatrix}$

Basically each sequential keys $y$ is a row in your input matrix. You can think of each matrix as some kind of image you feed to your LSTM network.
Hope that helps you any further.

| improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.