question: TFT for time series classification

[DominikKrock]

Hi,
I would like to use the temporal fusion transformer(TFT) for time-series-classification. For our illustrative example we have 2 target classes (True, False) for the cancellation of a membership.

Illustrative example:
Suppose you have an e-commerce shop, and you want to find out whether a customer will cancel his premium membership (like Amazon-prime) based on his shopping behavior.

Parameters for TimeSeriesDataSet:
target = [‘cancellation’]
group_ids= [‘customerID’]
static_categoricals = [‘zip’, ‘gender’]
time_varying_known_reals = [‘time_idx’]
time_varying_unknown_categoricals = [‘shopping_event’]
time_varying_unknown_reals = [‘age_at_shopping_event’]
max_prediction_length=1

Is it convenient to use the TFT for this kind of task?
How does the shape of the target vector look like, e.g.

• Method 1
  - always True for customers who cancel the membership?
  - always False for customers who don’t cancel the membership?
• Method 2
  - False until the customer cancels the membership. Then always True?

I have already implemented a BinaryCrossEntropyLoss metric analogous to the CrossEntropyLoss.
If you like, I can make a merge request for the BinaryCrossEntropyLoss.

Thanks for your ideas

[jdb78]

PRs are always very welcome!

I would generally go with method 2 because you do not want to use unknown information. Further, I would use max_prediction_length=1. You can use a label as well if someone will cancel their subscription in the next 3 months. Ensure that you are not using the cancellation target as feature. You might want to add a relative time index as well (see the docs how to add this automatically with the timeseriesdataset.

Effectively, you will this way use mostly the encoder of the TFT.

[jsolson4]

I have already implemented a BinaryCrossEntropyLoss metric analogous to the CrossEntropyLoss.
If you like, I can make a merge request for the BinaryCrossEntropyLoss.

Can you please share your implementation of BinaryCrossEntropyLoss?