Comparing the TreeSHAP and Exact explainers¶
We define the same dataset as exposed in gh-2345.
In [1]:
import numpy as np
X = np.vstack([
[[0, 0]] * 400,
[[0, 1]] * 100,
[[1, 0]] * 100,
[[1, 1]] * 400,
])
y = np.array(
[0] * 400 +
[50] * 100 +
[50] * 100 +
[100] * 400
)
Define two trees with different split ordering¶
We vary the random_state to make sure that
tree_1considersX[0]as a root splittree_2considersX[1]as a root split
Note that the 2 trees compute the same prediction on any data points. They are two distinct implementations of the same mathematical decision function.
In [2]:
import matplotlib.pyplot as plt
from sklearn.tree import DecisionTreeRegressor
from sklearn.tree import plot_tree
_, ax = plt.subplots(figsize=(10, 6))
tree_1 = DecisionTreeRegressor(random_state=2)
tree_1.fit(X, y)
_ = plot_tree(tree_1, ax=ax)
In [3]:
_, ax = plt.subplots(figsize=(10, 6))
tree_2 = DecisionTreeRegressor(random_state=0)
tree_2.fit(X, y)
_ = plot_tree(tree_2, ax=ax)
Explaining the prediction for a given data point with TreeSHAP¶
In [4]:
X_test = np.array([[1, 1]])
In [5]:
from shap import Explainer
In [6]:
tree_1_explainer = Explainer(model=tree_1, algorithm="tree")
tree_1_explainer(X_test)
Out[6]:
.values = array([[32.5, 17.5]]) .base_values = array([[50.]]) .data = array([[1, 1]])
In [7]:
tree_2_explainer = Explainer(model=tree_2, algorithm="tree")
tree_2_explainer(X_test)
Out[7]:
.values = array([[17.5, 32.5]]) .base_values = array([[50.]]) .data = array([[1, 1]])
We observe that we can reproduce the bug regarding the asymmetry of SHAP values reported in the original issue linked above.
Explaining the prediction for a given data point with the Exact explainer¶
In [8]:
from shap.explainers import Exact
from shap.maskers import Independent
explainer = Exact(tree_1.predict, Independent(X, max_samples=X.shape[0]))
explainer(X_test)
(array([1, 1]),)
Out[8]:
.values = array([[25., 25.]]) .base_values = array([50.]) .data = array([[1, 1]])
In [9]:
explainer = Exact(tree_2.predict, Independent(X, max_samples=X.shape[0]))
explainer(X_test)
(array([1, 1]),)
Out[9]:
.values = array([[25., 25.]]) .base_values = array([50.]) .data = array([[1, 1]])
The Exact explainer is not subject to the asymmetry problem.
We can as well check that a Python implementation of the same decision function leads to the same explaination.
In [11]:
def my_predict_one(x):
if x[0] < 0.5 and x[1] < 0.5:
return 0
elif x[0] > 0.5 and x[1] > 0.5:
return 100
else:
return 50
def my_predict(X):
return np.array(
[my_predict_one(x) for x in X]
)
explainer = Exact(my_predict, masker=Independent(X, max_samples=X.shape[0]))
explainer(X_test)
(array([1, 1]),)
Out[11]:
.values = array([[25., 25.]]) .base_values = array([50.]) .data = array([[1, 1]])
Checking the ACV implementation of SHAP values¶
In [23]:
from acv_explainers import ACVTree
ACVTree(tree_1, X).shap_values(X_test)
100%|████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 1216.45it/s]
Out[23]:
array([[[25.],
[25.]]])
In [19]:
ACVTree(tree_2, X).shap_values(X_test)
100%|████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 1452.82it/s]
Out[19]:
array([[[25.],
[25.]]])
Checking the FastTreeSHAP implementation¶
We now observe that this bug is also present in the FastTreeSHAP implementation.
In [11]:
%pip install -q fasttreeshap
ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
poetry 1.1.13 requires packaging<21.0,>=20.4, but you have packaging 21.3 which is incompatible.
Note: you may need to restart the kernel to use updated packages.
In [12]:
import fasttreeshap
fasttreeshap.TreeExplainer(tree_1, algorithm="v2")(X_test)
Out[12]:
.values = array([[32.5, 17.5]]) .base_values = array([[50.]]) .data = array([[1, 1]])
In [13]:
fasttreeshap.TreeExplainer(tree_2, algorithm="v2")(X_test)
Out[13]:
.values = array([[17.5, 32.5]]) .base_values = array([[50.]]) .data = array([[1, 1]])