『機械学習を解釈する技術』の紹介 / Devsumi2022

Transcript

1. 2022/2/17 Developers Summit 2022 @dropout009

2. TVISION INSIGHTS Twitter: @dropout009 Speaker Deck: dropout009 Blog: https://dropout009.hatenablog.com/

3. Deep Learning GBDT Random Forest

4. None

5. 画像認識、⾳声認識、機械翻訳、… Deep Learning, Gradient Boosting Decision Tree, Random Forest, ...

   pytorch, LightGBM, scikit-learn, ...

6. LM GBDT RF DL GLM

7. None

8. LM GBDT RF DL GLM +

9. None

10. 4 PFI PD + ICE SHAP

11. 1. PFI 2. PD 3. ICE 4. SHAP 住宅価格 =

    𝛽! + 𝛽" 部屋の数 + 𝛽# 駅からの距離 + ノイズ

12. Partial Dependence PD 𝑌 = 𝑋! − 5𝑋" + 10𝑋"𝑋#

    + 𝜖 𝑋!, 𝑋" ∼ Uniform −1, 1 𝑋# ∼ Bernoulli 0.5 𝜖 ∼ 𝒩(0, 0.01)

13. Python Partial Dependence @dataclass class PartialDependence: """Partial Dependence (PD)""" estimator:

    Any X: np.ndarray var_names: list[str] def _counterfactual_prediction( self, idx_to_replace: int, value_to_replace: float ) -> np.ndarray: """ある特徴量の値を置き換えたときの予測値を求める""" X_replaced = self.X.copy() X_replaced[:, idx_to_replace] = value_to_replace y_pred = self.estimator.predict(X_replaced) return y_pred ……… $ PD> 𝑥> = 1 𝑁 * ?@" A + 𝑓(𝑥> , 𝒙?,∖> )
14. None