Fixing the transformer behaviour on sparse features #122
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
…iner and Dataset.createTransformers to expose this behaviour.
…p more immutable as it should be.
…a provenance serialization issue in IDFTransformation.
Core/src/main/java/org/tribuo/transform/transformations/MeanStdDevTransformation.java
Show resolved
Hide resolved
Co-authored-by: Jack Sullivan <jack.t.sullivan@oracle.com>
There was a problem hiding this comment.
This has the changes that we discussed on Friday, but looking over it now, the implicit zeroes vs. densify behavior interactions aren't really explicitly characterized anywhere. I think it would be a good idea to have one place that clearly explains how all four possible settings work that can referred to for more detail in all the various constructors/methods where the interaction is important.
|
Ok, sure. I'll add something to the package-info in |
JackSullivan
reviewed
Apr 2, 2021
JackSullivan
reviewed
Apr 2, 2021
JackSullivan
reviewed
Apr 2, 2021
JackSullivan
reviewed
Apr 2, 2021
JackSullivan
requested changes
Apr 2, 2021
Co-authored-by: Jack Sullivan <jack.t.sullivan@oracle.com>
Co-authored-by: Jack Sullivan <jack.t.sullivan@oracle.com>
Co-authored-by: Jack Sullivan <jack.t.sullivan@oracle.com>
Co-authored-by: Jack Sullivan <jack.t.sullivan@oracle.com>
|
Thanks for the updates. |
JackSullivan
approved these changes
Apr 2, 2021
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Description
This PR adds implementations of
observeSparseto all the transformers (apart fromIDFTransformationwhich already had it). As a consequence to preserve the 4.0 behaviour the transformation fitting methods grow a new argument which turns on or off the use of theobserveSparsemethod. It also switchesDoubleFieldProcessorover so that it always emits values if they are parsable doubles. Before it would elide zero values, but this makes it much harder to implement transformations which should touch every value because they are numerical (rather than categorical encoded as a double).One downside of this implementation is that it's not possible to change the behaviour of
observeSparseon a per feature basis (due to it being a tricky breaking API change), which makes it difficult to apply an IDFTransformation to text features while simultaneously transforming features which should ignore the implicit zeros. If this behaviour is required then twoTransformationMaps can be applied in sequence to aDataset. We'll note this in the release notes for 4.1.Motivation
The observeSparse change is so the transformers actually have the documented behaviour, and then that rippled into adding new overloads for the transformation fitting methods. The
DoubleFieldProcessorchange is because it's tricky to correctly transform numerical data without it.