Generalize handling interaction between feature selectors and categorical features
- 主要言語
- Python
- スター
- 850
- フォーク
- 96
- PR マージ指標
- 30日以内にマージされた PR はありません
説明
In #3419 we added a fix to handle Email and URL features in default algorithm of such features by backtracking and looking at not only the feature provenance `EmailFeaturizer` and `URLFeaturizer` but also the feature provenance from `OneHotEncoder` for feature selection. This is because `EmailFeaturizer` and `URLFeaturizer` both generates categorical features that in turn get encoded by the `OneHotEncoder`. The flow is as follows:
`url` ---URLFeaturizer--> `URL_TO_DOMAIN(url)`----`OneHotEncoder`---->`URL_TO_DOMAIN(url)_alteryx.com`
#3419 introduces a fix specifically for Email and URL features but this exact issue could pop up again if another logical type is introduced that EvalML generates categorical features from.
Acceptance Criteria:
- generalize solution in #3419 to apply to future logical types where EvalML will create a categorical feature from
------------------------------------------------
"I think this is good but I feel like we need something more general here long term. If a new logical type is introduced, we might fall into the same bug again. Let's file an ice box issue to not forget. I can try to elaborate more there."
_Originally posted by @freddyaboulton in https://github.com/alteryx/evalml/pull/3419#discussion_r839739138_
コントリビューションガイド
評価
この issue はまだ評価されていません。