Status of MLDataPattern porting

[CarloLucibello]

A list of what is currently exported from MLDataPattern.jl.

TO PORT

• [x] getobs, getobs! and nobs. #1
  • nobs is now numobs;
  • obsdim argument is dropped from the interface
• [x] randobs #1
• [x] datasubset, DataSubset #4
• [x] shuffleobs #5
• [x] splitobs #5
• [x] DataView #5
  • [ ] Consider removal
• [x] obsview, ObsView #5
  • [ ] Consider removal #8
• [x] batchview, BatchView #6
• [x] batchsize #6
• [ ] slidingwindow, SlidingWindow
• [ ] stratifiedobs
• [x] oversample, undersample #10
• [x] kfolds #9
• [x] leaveout #9
• [x] eachobs #9
• [x] eachbatch #9

NOT TO BE PORTED

• BufferGetObs
• RandomObs, RandomBatches
• BalancedObs
• FoldView
• targets
• eachtarget

[CarloLucibello]

We can consider this essentially done

[rmkn85]

Hi, what about stratifiedobs and slidingwindow? Were they explicitly excluded on purpose? Thanks

[CarloLucibello]

not really, we just didn't port code that we weren't sure was going to be useful. I think stratifiedobs should go in, less sure of slidingwindow but didn't look much into it and alternatives in the ecosystem.

[rmkn85]

Just to clarify, I came here specifically for missing stratifiedobs. It is needed to replicate the behaviour of Python's sklearn.model_selection.train_test_split([...] stratify=true)

Asked about slidingwindow on the way, since it was the only other one unchecked but not in the list of explicitly "not to be ported", but I don't have any use-case for it.

[kpa28-git]

I use slidingwindow often for time series data. Haven't looked too much for a replacement but the closest I've found is IterTools.jl partition. It has a similar interface but returns a tuple iterator

[kpa28-git]

Also found DSP.Periodograms.arraysplit which is similar to slidingwindow but you set the overlap instead of the stride. So far slidingwindow is the fastest of the three because it returns views.