Variable interpretation

• explainer.excepted_value
  • The expectation of prediction results is sometimes the mean value of a batch of data prediction results??
  • Sub tags. If there are multiple categories, each category will have an expected value. When analyzing the shap value, select the excluded of the corresponding tag_ value
  • When interpreting the prediction results, starting from this value, each feature has an impact on the prediction results and finally determines the output of the model
  • This result should also be log odd, because it will have a negative number
• Log odd: log probability = log p / (1-p)
  • Assuming that the probability of a positive sample is 0.01, the logarithmic probability (based on 10) = log 0.01 / 0.99 = -3.9
  • Assuming that the probability of a positive sample is 0.99, the logarithmic probability (based on 10) = log 99 = 1.9
• shap_values
  • shape=(n_labels, n_rows, n_cols)
  • The information contained is that there is a value for each label, line and feature
  • In fact, when we analyze, we will look at the prediction analysis of each label separately. For the second category, we take the shap value with label 1, that is, shap_values[1]
  • When each row is analyzed separately, it is called shap_values[1].iloc[index]

Kernel explainer principle

To calculate one of the features of a record_ Values as an example to illustrate the calculation process:

1. Convert sample features into numbers
2. Randomly generate N random numbers as a mask, and calculate the original features with this mask to obtain n new features
3. The new n features are brought into the model prediction to obtain the prediction results. The change of the prediction results and the law of that feature are observed to obtain the shap_value

Example of interpreting prediction results using shap

Training lightgbm model:

import lightgbm as lgb
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)
model
# <lightgbm.basic.Booster object at 0x000001B96CBFAD48>

Build the interpreter and get the expected value:

explainer = shap.TreeExplainer(model)  #  Using the tree model interpreter
explainer
# Out[5]: <shap.explainers._tree.Tree at 0x1b96cca5e48>
expected_value = explainer.expected_value  
expected_value
# Out[7]: array([-2.43266725])

Build the interpreter and get the expected value:

features.shape
# Out[17]: (20, 12)
shap_values = explainer.shap_values(features)  # Calculate shap_values
len(shap_values)
# Out[12]: 2
shap_values[0].shape
# Out[13]: (20, 12)

Draw decision diagram:

shap.decision_plot(base_value=expected_value, shap_values=shap_values[1][:1], features=features_display[:1])

Output:

decision_plot explanation:

1. The x-axis is the output of the model, the y-axis is the feature name, and the line represents the prediction process
2. The output of the model is usually a probability, and the logarithmic probability is the output of the model
3. shap_values: the shap of the data to be interpreted_ values
4. Features: it is estimated that it is used to extract column names. You can not pass or use features_ Names instead
5. The greater the amplitude of characteristic pull, the greater the impact on the prediction results. The impact is divided into positive impact and negative impact
6. Find the features that have a great impact on the positive and their values to explain the prediction results

Efforts can also be used to explain the prediction results:

shap.force_plot(explainer.expected_value[0], shap_values[1][:1], features[:1], matplotlib=True)