search

py-why / EconML

ALICE (Automated Learning and Intelligence for Causation and Economics) is a Microsoft Research project aimed at applying Artificial Intelligence concepts to economic decision making. One of its goals is to build a toolkit that combines state-of-the-art machine learning techniques with econometrics in order to bring automation to complex causal inference problems. To date, the ALICE Python SDK (econml) implements orthogonal machine learning algorithms such as the double machine learning work of Chernozhukov et al. This toolkit is designed to measure the causal effect of some treatment variable(s) t on an outcome variable y, controlling for a set of features x.
https://www.microsoft.com/en-us/research/project/alice/
Other
3.76k stars 713 forks source link

Reproducible error: SHAP ExplainerError: Additivity check failed in TreeExplainer #873

Open yanisvdc opened 5 months ago

yanisvdc commented 5 months ago

This is related to issue #866 

import pandas as pd
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import MinMaxScaler

# Create a sample DataFrame X
np.random.seed(0)  # for reproducibility
data = {
    'Category': np.random.randint(2, size=100),
    'Value': np.random.randint(100, size=100)
}
X = pd.DataFrame(data)

# Creating binary arrays T and y
T = np.random.randint(2, size=100)
y = np.random.randint(2, size=100)

# Splitting the data into training and testing sets
X_train, X_test, T_train, T_test, y_train, y_test = train_test_split(X, T, y, test_size=0.2, random_state=42)

to_scale = ["Value"]
scaler = MinMaxScaler()
# Fit scaler on training data and transform both training and testing data
X_train[to_scale] = scaler.fit_transform(X_train[to_scale])
X_test[to_scale] = scaler.transform(X_test[to_scale])

est = CausalForestDML(model_y=RandomForestClassifier(), model_t=DummyClassifier(strategy='uniform'), random_state=123, discrete_outcome=True,discrete_treatment=True)
est.fit(y_train, T_train, X=X_train)
shap_values = est.shap_values(X_train[:20])
shap.plots.beeswarm(shap_values['Y0']['T0_1'])

If you try to comment out or remove the lines:

to_scale = ["Value"]
scaler = MinMaxScaler()
# Fit scaler on training data and transform both training and testing data
X_train[to_scale] = scaler.fit_transform(X_train[to_scale])
X_test[to_scale] = scaler.transform(X_test[to_scale])

the code runs correctly.

It seems that the scaler induces an issue, it is the same with StandardScaler(). My guess is that float rounding errors are responsible for this behavior. I believe that there is a PR to add check_additivity = False, which seems to be the only way to resolve this, unless I am missing something.

jcreinhold commented 5 months ago

Thanks for the code. I was trying to write a consistent unit test for PR #872 (which fixes this issue and #866); however, I'm unable to reproduce the error with the following code (structured as a unit test in test_shap.py):

def test_check_additivity_handling(self):
    np.random.seed(0)
    n = 100
    X = np.hstack([np.random.randint(2, size=(n, 1)), np.random.randint(100, size=(n, 1))])
    T = np.random.randint(2, size=(n,))
    Y = np.random.randint(2, size=(n,))
    X_train, _, T_train, _, Y_train, _ = train_test_split(X, T, Y, test_size=0.2, random_state=42)
    scaler = MinMaxScaler()
    X_train[:, 1:2] = scaler.fit_transform(X_train[:, 1:2])
    est = CausalForestDML(
        model_y=RandomForestClassifier(),
        model_t=DummyClassifier(strategy="uniform"),
        random_state=123,
        discrete_outcome=True,
        discrete_treatment=True
    )
    est.fit(Y_train, T_train, X=X_train)
    shap_values = est.shap_values(X_train[:20])
    assert shap_values is not None

As discussed in my second comment in #866, I don't believe there's a consistent way to reproduce this error across machines—for various reasons. However, #872 will fix this issue regardless.

jcreinhold commented 5 months ago

@yanisvdc Do you have the literal values for X_train, T_train, and y_train that create the error (after scaling)? I can try hard-coding those values in and seeing if that reproduces the error. Given they're size 80, this seems feasible.

Could you please try the following (on the set of variables that cause the error):

arrays = {
    "X": X_train,  # after scaling (potentially call .to_numpy() on it if it's a df)
    "T": T_train.squeeze(),
    "Y": y_train.squeeze(),
}
for name, arr in arrays.items():
    mlw = 20 if name != "X" else 50
    arr_str = np.array2string(arr, separator=", ", floatmode="maxprec", precision=16, max_line_width=mlw)
    print(f"{name}:\n\n{arr_str}")

and copy the results here in a markdown codeblock for each variable?

yanisvdc commented 5 months ago

Thanks for your comments. Your test named test_check_additivity_handling() should fail by adding "astype(float)" to X. However, your test passes by removing it. The problem is when you remove "astype(float)" to X, the MinMaxScaler outputs zeros as it casts the values to int to match the dtype. Thus, your current test passes but that is only because X_train ends up being an array with binary integer values. In the more general case, with "astype(float)" added to X, MinMaxScaler outputs floats between 0 and 1, and the error displayed below will be raised (see after the code containing "astype(float)" added to X)

np.random.seed(0)
n = 100
X = np.hstack([np.random.randint(2, size=(n, 1)), np.random.randint(100, size=(n, 1))]).astype(float)
T = np.random.randint(2, size=(n,))
Y = np.random.randint(2, size=(n,))
X_train, _, T_train, _, Y_train, _ = train_test_split(X, T, Y, test_size=0.2, random_state=42)
scaler = MinMaxScaler()
X_train[:, 1:2] = scaler.fit_transform(X_train[:, 1:2])
est = CausalForestDML(
    model_y=RandomForestClassifier(),
    model_t=DummyClassifier(strategy="uniform"),
    random_state=123,
    discrete_outcome=True,
    discrete_treatment=True
)
est.fit(Y_train, T_train, X=X_train)
shap_values = est.shap_values(X_train[:20])
shap.plots.beeswarm(shap_values['Y0']['T0_1'])

Error message:

ExplainerError: Additivity check failed in TreeExplainer! Please ensure the data matrix you passed to the explainer is the same shape that the model was trained on. If your data shape is correct then please report this on GitHub. This check failed because for one of the samples the sum of the SHAP values was -0.061160, while the model output was -0.059755. If this difference is acceptable you can set check_additivity=False to disable this check.

Please let me know if you experience the same behavior.

yanisvdc commented 5 months ago

@yanisvdc Do you have the literal values for X_train, T_train, and y_train that create the error (after scaling)? I can try hard-coding those values in and seeing if that reproduces the error. Given they're size 80, this seems feasible.

Could you please try the following (on the set of variables that cause the error):

arrays = {
    "X": X_train,  # after scaling (potentially call .to_numpy() on it if it's a df)
    "T": T_train.squeeze(),
    "Y": y_train.squeeze(),
}
for name, arr in arrays.items():
    mlw = 20 if name != "X" else 50
    arr_str = np.array2string(arr, separator=", ", floatmode="maxprec", precision=16, max_line_width=mlw)
    print(f"{name}:\n\n{arr_str}")

and copy the results here in a markdown codeblock for each variable?

To answer this, here are the exact values:

X:

[[0.                , 0.202020202020202 ],
 [1.                , 0.4040404040404041],
 [1.                , 0.696969696969697 ],
 [0.                , 0.4343434343434344],
 [0.                , 0.5151515151515152],
 [0.                , 0.2323232323232323],
 [0.                , 0.                ],
 [0.                , 0.6767676767676768],
 [1.                , 0.5252525252525253],
 [0.                , 0.5454545454545455],
 [0.                , 0.1515151515151515],
 [1.                , 0.98989898989899  ],
 [1.                , 0.9494949494949496],
 [1.                , 0.                ],
 [0.                , 0.2626262626262627],
 [1.                , 0.686868686868687 ],
 [0.                , 0.4343434343434344],
 [0.                , 0.8686868686868687],
 [1.                , 0.98989898989899  ],
 [0.                , 0.797979797979798 ],
 [1.                , 0.3535353535353536],
 [1.                , 0.0303030303030303],
 [1.                , 0.0303030303030303],
 [0.                , 0.5252525252525253],
 [1.                , 0.9494949494949496],
 [1.                , 0.4848484848484849],
 [1.                , 0.303030303030303 ],
 [1.                , 0.9595959595959597],
 [1.                , 0.6565656565656566],
 [1.                , 1.                ],
 [0.                , 0.6464646464646465],
 [0.                , 0.686868686868687 ],
 [0.                , 0.1313131313131313],
 [1.                , 0.2121212121212121],
 [1.                , 0.7676767676767677],
 [1.                , 0.5656565656565657],
 [1.                , 0.4747474747474748],
 [1.                , 0.101010101010101 ],
 [0.                , 0.4242424242424243],
 [0.                , 0.7272727272727273],
 [1.                , 0.8181818181818182],
 [0.                , 0.9696969696969697],
 [1.                , 1.                ],
 [1.                , 0.5858585858585859],
 [0.                , 0.9494949494949496],
 [0.                , 0.1111111111111111],
 [1.                , 0.0202020202020202],
 [0.                , 0.0202020202020202],
 [0.                , 0.5858585858585859],
 [1.                , 0.6161616161616162],
 [0.                , 0.6060606060606061],
 [1.                , 0.101010101010101 ],
 [0.                , 0.2727272727272728],
 [0.                , 0.6262626262626263],
 [1.                , 0.1414141414141414],
 [0.                , 0.5050505050505051],
 [1.                , 0.1919191919191919],
 [1.                , 0.4848484848484849],
 [1.                , 0.6666666666666667],
 [1.                , 0.1414141414141414],
 [0.                , 0.3636363636363636],
 [1.                , 0.0303030303030303],
 [1.                , 0.7777777777777778],
 [0.                , 0.5050505050505051],
 [1.                , 0.8484848484848485],
 [1.                , 0.8282828282828284],
 [1.                , 0.696969696969697 ],
 [1.                , 0.7575757575757577],
 [0.                , 0.3535353535353536],
 [1.                , 0.1313131313131313],
 [1.                , 0.9595959595959597],
 [0.                , 0.3838383838383839],
 [1.                , 0.98989898989899  ],
 [0.                , 0.2424242424242424],
 [0.                , 0.494949494949495 ],
 [1.                , 0.4141414141414142],
 [1.                , 0.3232323232323233],
 [0.                , 0.5858585858585859],
 [0.                , 0.7272727272727273],
 [1.                , 0.6767676767676768]]
T:

[0, 1, 0, 0, 1, 0,
 0, 1, 1, 0, 0, 1,
 1, 1, 1, 0, 1, 0,
 1, 1, 1, 1, 1, 1,
 0, 1, 0, 0, 0, 1,
 1, 1, 0, 1, 1, 1,
 1, 0, 0, 0, 1, 1,
 1, 1, 1, 1, 1, 1,
 0, 1, 0, 0, 1, 1,
 1, 0, 0, 1, 1, 1,
 1, 1, 0, 1, 1, 0,
 1, 1, 1, 1, 1, 0,
 1, 0, 0, 0, 1, 1,
 1, 0]
Y:

[1, 1, 1, 0, 1, 1,
 0, 1, 1, 1, 1, 1,
 1, 1, 1, 1, 1, 0,
 1, 0, 0, 0, 1, 0,
 0, 0, 0, 0, 1, 1,
 0, 0, 0, 0, 0, 0,
 1, 0, 1, 1, 0, 0,
 1, 0, 1, 1, 1, 0,
 0, 1, 1, 1, 1, 1,
 1, 1, 0, 0, 1, 0,
 1, 1, 1, 0, 0, 0,
 1, 0, 1, 0, 0, 1,
 1, 1, 1, 1, 1, 1,
 0, 0]