scikit-learn/scikit-learn

Tolerance on RFECV

Open

#7,559 建立於 2016年10月3日

在 GitHub 查看
 (14 留言) (0 反應) (0 負責人)Python (66,084 star) (27,020 fork)batch import
EnhancementModeratehelp wantedmodule:feature_selection

描述

Hi,

Is there any way to specify a tolerance when determining the number of optimal features when using RFECV (like n the CARET package)? Currently when using RFECV with many of the tree-based classifiers, the removal of the least important n features often results in a very slight reduction in model accuracy. This means that all of the features end up being selected as 'optimal', even though the reduction in model accuracy is very slight if we used a much smaller set of features. A tolerance setting, like 1% would allow only the features to be selected that would otherwise cause a large drop in model accuracy.

Steve

貢獻者指南

技術棧
python
領域
machine learningdata
議題類型
feature
難度面向新貢獻者的預計實作難度，1 表示很小改動，5 表示專家級工作。
3
預計時間有經驗貢獻者完成調查、實作、測試並準備 pull request 的粗略時間範圍。
half day
活動狀態議題目前的可參與程度：新鮮、活躍、陳舊、阻塞或等待維護者輸入。
stale
清晰度議題是否清楚說明預期改動、驗收標準和下一步。
mostly clear
前置要求
Understanding of RFECVBasic knowledge of feature selectionFamiliarity with cross validation
新手友善度1-100 的估計分數，表示該議題對首次貢獻者的友善程度。
70
研究方向
Investigate the current RFECV implementation in sklearn/feature selection/ rfe.py and the associated tests. Review the discussion in issue #7559 for proposed tolerance approaches. Consider looking at how CARET's rfe function implements a tolerance parameter. The goal is to add an optional `tolerance` parameter to RFECV that stops feature elimination when the cross validation score does not improve beyond the tolerance.