In [2]:
from pprint import pprint
import lightgbm as lgb
import matplotlib.pyplot as plt
import numpy as np
import pickle
import shap
from sklearn.model_selection import train_test_split, StratifiedKFold
import warnings

X, y = shap.datasets.adult()
X_display, y_display = shap.datasets.adult(display=True)

# create a train/test split
random_state = 7
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=random_state)
d_train = lgb.Dataset(X_train, label=y_train)
d_test = lgb.Dataset(X_test, label=y_test)

params = {
    "max_bin": 512,
    "learning_rate": 0.05,
    "boosting_type": "gbdt",
    "objective": "binary",
    "metric": "binary_logloss",
    "num_leaves": 10,
    "verbose": -1,
    "min_data": 100,
    "boost_from_average": True,
    "random_state": random_state
}

model = lgb.train(params, d_train, 10000, valid_sets=[d_test], early_stopping_rounds=50, verbose_eval=1000)

Training until validation scores don't improve for 50 rounds
Early stopping, best iteration is:
[644]	valid_0's binary_logloss: 0.278029
In [3]:
explainer = shap.TreeExplainer(model)
expected_value = explainer.expected_value
if isinstance(expected_value, list):
    expected_value = expected_value[1]
print(f"Explainer expected value: {expected_value}")

select = range(20)
features = X_test.iloc[select]
features_display = X_display.loc[features.index]

with warnings.catch_warnings():
    warnings.simplefilter("ignore")
    shap_values = explainer.shap_values(features)[1]
    shap_interaction_values = explainer.shap_interaction_values(features)
if isinstance(shap_interaction_values, list):
    shap_interaction_values = shap_interaction_values[1]

Explainer expected value: [-2.43266725]

Basic decision plot features

Refer to the decision plot of the 20 test observations below. Note: This plot isn't informative by itself; we use it only to illustrate the primary concepts.

  • The x-axis represents the model's output. In this case, the units are log odds.
  • The plot is centered on the x-axis at explainer.expected_value. All SHAP values are relative to the model's expected value like a linear model's effects are relative to the intercept.
  • The y-axis lists the model's features. By default, the features are ordered by descending importance. The importance is calculated over the observations plotted. This is usually different than the importance ordering for the entire dataset. In addition to feature importance ordering, the decision plot also supports hierarchical cluster feature ordering and user-defined feature ordering.
  • Each observation's prediction is represented by a colored line. At the top of the plot, each line strikes the x-axis at its corresponding observation's predicted value. This value determines the color of the line on a spectrum.
  • Moving from the bottom of the plot to the top, SHAP values for each feature are added to the model's base value. This shows how each feature contributes to the overall prediction.
  • At the bottom of the plot, the observations converge at explainer.expected_value.
In [4]:
shap.decision_plot(expected_value, shap_values, features_display)

Like the force plot, the decision plot supports link='logit' to transform log odds to probabilities.

In [5]:
shap.decision_plot(expected_value, shap_values, features_display, link='logit')
In [6]:
# Our naive cutoff point is zero log odds (probability 0.5).
y_pred = (shap_values.sum(1) + expected_value) > 0
misclassified = y_pred != y_test[select]
shap.decision_plot(expected_value, shap_values, features_display, link='logit', highlight=misclassified)

Let's inspect the misclassified observation by plotting it alone. When a single observation is plotted, its corresponding feature values are displayed. Notice that the shape of the line has changed. Why? The feature order has changed on the y-axis based on the feature importance for this lone observation. The section "Preserving order and scale between plots" shows how to use the same feature order for multiple plots.

In [7]:
shap.decision_plot(expected_value, shap_values[misclassified], features_display[misclassified],
                   link='logit', highlight=0)
In [8]:
shap.decision_plot(expected_value, shap_interaction_values, features, link='logit', 
                   feature_order='hclust', feature_display_range=slice(None, None, -1))
In [9]:
p = 0.4  # Probability 0.4
new_base_value = np.log(p / (1 - p))  # the logit function
shap.decision_plot(expected_value, shap_values, features_display, link='logit', new_base_value=new_base_value)
In [ ]: