Learnable Bin Size Parameter

Hello all !

I’m working on time series prediction with sparse signal, meaning most of the response variable are 0’s.
Usually, we dealt with this using binning and the bin size would be computed manually.

Lately we’ve been experimenting with deep learning and I was wondering if there would be a way to include a learnable parameter that could “learn” to select the right bin size given the input…

I was inspired by concrete dropout where the drop out rate is learned but since a bin size would actually change the input data i’m a little bit confused about the training process and back-prop.

Has anyone dealt with something like this before ?


I’m not an expert, but to use backprop, your objective function would have to be differentiable. Bin-size is a non-differentiable objective. You could learn on the timeseries itself using an RNN, CNN, or Transformer. Otherwise, you either have to find a relaxation of your objective (ie make it differentiable) or use a learning algorithm such as evolutionary methods which don’t require differentiable objectives.

That makes a lot of sens, thanks you very much for your answer. I see i still need to think of a proper solution.