Encoding sparse timeseries to input vectors for deep NNs

How would you model the following input data in deep learning setting.

Each user generates events (typically tens of events, not hundreds per user) over, say, 30 days. We want to use these events as an input to a deep learning neural network for a classification task.

One option is to split 30 days to e.g. 1-hour slots and encode event type to a value of a slot. Most slots will be empty, but at least in theory 1D convolution could be used on this kind of input.

Are there better ways to encode sparse timeseries to input for deep learning network?

If it’s sparse but periodic then an FFT transformed representation might make sense here.