I am trying to solve a multimodal regression task. I have two types of time series available. 

  • Continuous data
  • One-hot encoded event data

Both time series are sparsely and irregularly sampled, with continous timestamps. For each point in time, a target label is available. I would like to find a way to create a joined embedding of both data types, ideally regularly sampled, so it can be used as input for a standard RNN-block.

So far, I was not able to find a good approach for this problem.


1 Answer 1


More information is needed, e.g., samples of both data types would help.

-Is the meaning of the target variable(s) the same for both?

-What are the unit(s) of the timestamps (hourly, daily, weekly...)?

-Is the one-hot encoding based on the same information as the continuous data, if not, how do they differ?

One approach to equilibrating disparate time series is to aggregate by timestamp and merge based on time. This means taking disaggregated and irregular time series and aggregating them to a higher level, e.g., hourly time periods can be rolled up to daily, daily to weekly, etc.

Depending on how irregular the time series are, this may or may not regularize the information.

But in the absence of more information, this is speculation.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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