You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I am often using random datasets (typically with make_classification). However I often find myself having to add more realistic features to the dataset:
missing data, sometime just to test the pipeline (missing at random would be fine), or sometimes to look for more complex phenomenons (missingnes not at random, possibly depending on the target)
categorical: categoricals variables often need to be handled specifically. I usually introduce categoricals with binning a continuous value, then transforming to strings.
It would be nice to have both of those in datasets generation.
Describe your proposed solution
Introduce parameters to allow for generation of missing data (proportion of missingness, type of missingness - at random, not at random).
Introduce parameters to allow for generation of categorical features (number of features, type of repartition in categories - even - uneven - pareto.
Describe alternatives you've considered, if relevant
I usually handle this by hand.
Additional context
Could be used to illustrate imputing techniques, encoding techniques.
The text was updated successfully, but these errors were encountered:
@lcrmorin This would be great for testing! I would also suggest adding infinities as possible values, bcs they also break stuff quite often. Also, if randomly generated, making sure to always include at least one NaN and inf value
@lcrmorin I suggest adding a noise function or something similar which can generate structured randomness so as to make some sense in data and not pseudo-randomness. Perhaps something like Perlin Noise?
Regarding the missing values I recall the following issues/PRs: #6284 / #7084. It seems that the consensus was to have something similar to the ampute R package.
I almost a similar discussion for categorical features but I could not find. For sure, it would be handy to have those two parameters even though we could limit the complexity (e.g. only have a single missingness pattern)
Describe the workflow you want to enable
I am often using random datasets (typically with make_classification). However I often find myself having to add more realistic features to the dataset:
It would be nice to have both of those in datasets generation.
Describe your proposed solution
Introduce parameters to allow for generation of missing data (proportion of missingness, type of missingness - at random, not at random).
Introduce parameters to allow for generation of categorical features (number of features, type of repartition in categories - even - uneven - pareto.
Describe alternatives you've considered, if relevant
I usually handle this by hand.
Additional context
Could be used to illustrate imputing techniques, encoding techniques.
The text was updated successfully, but these errors were encountered: