How to explain prediction for a data with just a few features (from all features of training dataset)?

[williamty]

I have generated lightGBM models for prediction. I can explain the predictions with all features by filling user input data with NAs. Is there any way to explain prediction for the original user input data without filling it?

[apoplexi24]

You can run the LIME Explainer on select few columns/features too.
Change the LIME explainer from classification to regression if your model is used for regression.

// write logic to select the features that you want to run LIME Explainer on
// for example
selected_data = data[['col1', 'col2']]

// Create a LimeTabularExplainer
explainer = LimeTabularExplainer(selected_data.values, feature_names=selected_data.columns.values, mode="classification")

// select instance to explain
data_row = selected_data.iloc[0] # Get the first row in the selected data
// the num_features helps us select the features that we want to predict
explanation = explainer.explain_instance(data_row, lgbm_model.predict, num_features=len(selected_data.columns))

// Display the explanation
explanation.show_in_notebook()

The above method works for explaining the predictions when you want to have selective features. But if you want to generate predictions, your testing data (x_test) has to have the same features as the training data (x_train) in the ML model, or else it'll throw error of features not being the same.

[williamty]

@apoplexi24 Thank you for your kind reply!! It worked! By the way, I have also changed the code of predict function, setting the 'predict_disable_shape_check' parameter to true:
`
def predict_fn(x):
if len(np.array(x).shape) == 1:
# Reshape individual data points to 2D
return ldl.predict(np.array(x).reshape(1, -1), predict_disable_shape_check=True)
else:
# Predict for the entire dataset
return ldl.predict(x, predict_disable_shape_check=True)

def predict_fn_binary(x):
return np.column_stack((1 - predict_fn(x), predict_fn(x)))
`