3

Suppose I want to build a timeseries where each timestep is represented by a categorical array: the encoded sequences look like `[[2, 0, 5],[3, 1, 4],..]`

and each entry has a different number of possible values (categories).
For example the first entry has 0-3 values, the second 0-1 and so on...

I want to train an LSTM model in order to predict **the next timestep**. So I defined a one hot encoding of each entry by means of the maximum number of classes:

For example `[2, 0, 5]`

becomes

```
[[0. 0. 1. 0. 0. 0.],
[1. 0. 0. 0. 0. 0.],
[0. 0. 0. 0. 0. 1.]]
```

Unfortunately this kind of representation raises the error

ValueError: Invalid shape for y: (1, 3, 5)

I have three questions:

- Is it possible to pass a
**3d y target**to Keras? - Should I define a
**single one-hot encoding**which combines all the possible triplets of categorical values instead? The problem is that in this case I would lose the correlation between the occurrences of the labels in the same category, because each possible combination of labels would become independent from the other ones. - Should I
**only one-hot encode the target y**or also the input X?