tree influence: 決定木ベースのアルゴリズムに向けて、事例型説明（予測に影響した訓練データを出す方法）の手法をまとめたパッケージ
❌️ 途中で ×木の数のnp.arrayを作るタイミングがあり、例えば100万レコード×5万の決定木ではfloat32だとしても186GiBになってしまう
❌️ 2024年ごろから更新が停止しておりPython3.11以降はサポートしていない
import numpy as np
from sklearn.datasets import load_iris
from sklearn.model_selection import train_test_split
from lightgbm import LGBMClassifier
from tree_influence.explainers import BoostIn
# load iris data
data = load_iris()
X, y = data['data'], data['target']
# use two classes, then split into train and test
idxs = np.where(y != 2)[0]
X, y = X[idxs], y[idxs]
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.1, random_state=1)
# train GBDT model
model = LGBMClassifier().fit(X_train, y_train)
# fit influence estimator
explainer = BoostIn().fit(model, X_train, y_train)
# estimate training influences on each test instance
influence = explainer.get_local_influence(X_test, y_test) # shape=(no. train, no. test)
# extract influence values for the first test instance
values = influence[:, 0] # shape=(no. train,)
# sort training examples from:
# - most positively influential (decreases loss of the test instance the most), to
# - most negatively influential (increases loss of the test instance the most)
training_idxs = np.argsort(values)[::-1]