Model interpretability (SHAP, LIME) in ML Python - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

Understanding SHAP values

Which statement best describes what SHAP values represent in model interpretability?

A SHAP values show how much each feature contributes to pushing the model output from the average prediction to the current prediction.

B SHAP values are the raw input features used to train the model.

C SHAP values are the predicted probabilities of the model for each class.

D SHAP values are the weights of the neurons in a neural network.

💡 Hint

Think about how SHAP explains individual predictions by comparing to a baseline.

SHAP values quantify the contribution of each feature to the difference between the model's average output and the specific prediction, helping to explain why the model made that prediction.

❓ Predict Output

intermediate

Output of LIME explanation code

What will be the output of the following Python code snippet using LIME for a binary classification model?

ML Python

import numpy as np
from sklearn.linear_model import LogisticRegression
from lime.lime_tabular import LimeTabularExplainer

X_train = np.array([[0,1],[1,1],[0,0],[1,0]])
y_train = np.array([0,1,0,1])
model = LogisticRegression().fit(X_train, y_train)

explainer = LimeTabularExplainer(X_train, feature_names=['f1','f2'], class_names=['class0','class1'], discretize_continuous=False)
exp = explainer.explain_instance(np.array([1,0]), model.predict_proba, num_features=2)
print(exp.as_list())

A [('f2 <= 0.5', 0.5), ('f1 > 0.5', 0.3)]

B [('f1', 0.5), ('f2', -0.1)]

C [('f1', 0.3), ('f2', 0.2)]

D [('f1 = 1', 0.4), ('f2 = 0', 0.6)]

💡 Hint

LIME returns feature contributions as tuples with feature name and weight.

LIME explains the prediction by showing feature contributions with weights. For input [1,0], 'f1' contributes positively around 0.5 and 'f2' slightly negatively.

❓ Model Choice

advanced

Choosing between SHAP and LIME

You want to explain predictions of a complex black-box model on tabular data with many features. Which method is best if you need consistent global and local explanations?

A Use LIME because it provides consistent global explanations for all features.

B Use LIME because it is faster and does not require retraining the model.

C Use SHAP because it provides consistent additive feature attributions for both local and global explanations.

D Use neither; use partial dependence plots instead.

💡 Hint

Consider which method has a solid theoretical foundation for consistent explanations.

SHAP values have a strong theoretical basis ensuring consistent additive explanations that can be aggregated for global insights, unlike LIME which is local and can vary.

❓ Hyperparameter

advanced

Effect of number of samples in LIME

In LIME, increasing the 'num_samples' parameter during explanation will most likely:

A Decrease explanation accuracy and increase speed.

B Cause the model to retrain on more data.

C Have no effect on explanation quality or speed.

D Increase explanation accuracy but increase computation time.

💡 Hint

Think about how sampling affects the local surrogate model fit.

More samples allow LIME to better approximate the local decision boundary, improving explanation accuracy but requiring more computation time.

🔧 Debug

expert

Debugging SHAP value computation error

You run SHAP's TreeExplainer on a scikit-learn RandomForestClassifier but get the error: 'ValueError: Model output type not supported'. What is the most likely cause?

A You used TreeExplainer on a model that is not tree-based.

B You passed a multi-output regression model instead of classification.

C You passed raw model predictions instead of probabilities to SHAP.

D You did not convert the input data to a numpy array.

💡 Hint

Check if the model type matches the explainer requirements.

TreeExplainer only supports tree-based models like RandomForest or XGBoost. Using it on non-tree models causes this error.