Note
Go to the end to download the full example code.
ShaRP for classification on large datasets with mixed data types¶
This example showcases a more complex setting, where we will develop and interpret a classification model using a larger dataset with both categorical and continuous features.
sharp is designed to operate over the unprocessed input space, to ensure every
“Frankenstein” point generated to compute feature contributions are plausible. This means
that the function producing the scores (or class predictions) should take as input the
raw dataset, and every preprocessing step leading to the black box predictions/scores
should be included within it.
We will start by downloading the German Credit dataset.
import pandas as pd
import matplotlib.pyplot as plt
from sklearn.datasets import fetch_openml
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import OneHotEncoder, MinMaxScaler
from sklearn.compose import ColumnTransformer
from sklearn.linear_model import LogisticRegression
from sklearn.pipeline import make_pipeline
from sharp import ShaRP
df = fetch_openml(data_id=31, parser="auto")["frame"]
df.head(5)
Split X and y (input and target) from df and split train and test:
X = df.drop(columns="class")
y = df["class"]
categorical_features = X.dtypes.apply(
lambda dtype: isinstance(dtype, pd.CategoricalDtype)
).values
X_train, X_test, y_train, y_test = train_test_split(
X, y, test_size=0.1, random_state=42
)
Now we will set up model. Here, we will use a pipeline to combine all the preprocessing
steps. However, to use sharp, it is also sufficient to pass any function
(containing all the preprocessing steps) that takes a numpy array as input and outputs
the model’s predictions.
transformer = ColumnTransformer(
transformers=[
("onehot", OneHotEncoder(sparse_output=False), categorical_features),
("minmax", MinMaxScaler(), ~categorical_features),
],
remainder="passthrough",
n_jobs=-1,
)
classifier = LogisticRegression(random_state=42)
model = make_pipeline(transformer, classifier)
model.fit(X_train.values, y_train.values)
We can now use sharp to explain our model’s predictions! If we consider the dataset
to be too large, we have a few options to reduce computational complexity, such as
configuring the n_jobs parameter, setting a value on sample_size, or setting
measure=unary.
xai = ShaRP(
qoi="flip",
target_function=model.predict,
measure="unary",
sample_size=None,
random_state=42,
n_jobs=-1,
verbose=1,
)
xai.fit(X_test)
unary_values = pd.DataFrame(xai.all(X_test), columns=X.columns)
unary_values
0%| | 0/100 [00:00<?, ?it/s]
1%| | 1/100 [00:00<00:19, 5.14it/s]
3%|▎ | 3/100 [00:00<00:08, 11.01it/s]
6%|▌ | 6/100 [00:00<00:05, 17.90it/s]
10%|█ | 10/100 [00:00<00:04, 19.69it/s]
14%|█▍ | 14/100 [00:00<00:04, 19.88it/s]
18%|█▊ | 18/100 [00:00<00:04, 19.57it/s]
22%|██▏ | 22/100 [00:01<00:03, 20.32it/s]
26%|██▌ | 26/100 [00:01<00:03, 20.29it/s]
34%|███▍ | 34/100 [00:01<00:02, 26.29it/s]
42%|████▏ | 42/100 [00:01<00:01, 29.62it/s]
50%|█████ | 50/100 [00:02<00:01, 32.40it/s]
58%|█████▊ | 58/100 [00:02<00:01, 34.13it/s]
66%|██████▌ | 66/100 [00:02<00:01, 33.73it/s]
74%|███████▍ | 74/100 [00:02<00:00, 35.09it/s]
82%|████████▏ | 82/100 [00:02<00:00, 35.84it/s]
90%|█████████ | 90/100 [00:03<00:00, 36.60it/s]
97%|█████████▋| 97/100 [00:03<00:00, 41.34it/s]
100%|██████████| 100/100 [00:03<00:00, 30.56it/s]
Finally, we can plot the mean contributions of each feature:
plt.style.use("seaborn-v0_8-whitegrid")
fig, ax = plt.subplots()
xai.plot.bar(unary_values.mean(), ax=ax)
ax.set_ylim(bottom=0)
ax.tick_params(labelrotation=90)
fig.tight_layout()
plt.show()
Total running time of the script: (0 minutes 6.575 seconds)