eli5: Explain ML Predictions

Common errors

• ImportError: cannot import name 'load_boston' from 'sklearn.datasets'

  cause The `load_boston` dataset was deprecated in scikit-learn 1.0 and removed in scikit-learn 1.2. Your scikit-learn version is too new for this dataset.
  fix Update your code to use a modern dataset such as `fetch_california_housing` or `load_diabetes`.

• ValueError: Feature names passed were not known to the model.

  cause `eli5` could not determine the feature names, either because the model was trained without them (e.g., on numpy arrays instead of pandas DataFrames) or `feature_names` were not explicitly provided to the `eli5` explanation function.
  fix Ensure your training data is a Pandas DataFrame or explicitly pass a list of `feature_names` to `eli5.show_weights` or `eli5.show_prediction`, e.g., `eli5.show_weights(model, feature_names=X.columns.tolist())`.

• ImportError: cannot import name 'explain_weights' from 'eli5.sklearn'

  cause You are attempting to use an old API. Prior to `eli5` 0.10.0, explanations were typically done through backend-specific functions like `eli5.sklearn.explain_weights`. These have been unified into generic `eli5.show_weights` and `eli5.show_prediction`.
  fix Update your code to use the generic functions: replace `eli5.sklearn.explain_weights(model)` with `eli5.show_weights(model)` (and similarly for `explain_prediction`).

Warnings

• breaking The old backend-specific explanation functions like `eli5.sklearn.explain_weights` and `eli5.sklearn.explain_prediction` were removed and replaced by generic `eli5.show_weights` and `eli5.show_prediction`.
  Fix: Update your code to use `eli5.show_weights(...)` for global explanations and `eli5.show_prediction(...)` for individual predictions.
• gotcha `eli5.show_weights` and `eli5.show_prediction` return `IPython.display.HTML` objects in Jupyter notebooks for rich display. In a standard Python console, a string representation (often HTML source) is printed. If you need programmatic access (e.g., as a DataFrame or plain text), you must explicitly use formatters.
  Fix: For DataFrame output: `from eli5.formatters import as_dataframe; df = as_dataframe(eli5.show_weights(...))`. For plain text: `from eli5.formatters import as_text; text = as_text(eli5.show_weights(...))`.
• gotcha Many `scikit-learn` datasets (e.g., `load_boston`) are deprecated or removed in newer `scikit-learn` versions (1.2+). Code relying on these datasets will cause `ImportError` or `FutureWarning`.
  Fix: Use modern datasets like `fetch_california_housing` or `load_diabetes` (the non-deprecated version) for compatibility.

Install

• pip install eli5 Basic Installation
• pip install 'eli5[sklearn,lightgbm,xgboost,keras,nltk]' # Install with common extras Installation with common ML backends

Imports

• show_weights

  import eli5
  eli5.show_weights(...)

• show_prediction

  import eli5
  eli5.show_prediction(...)

• PermutationImportance

  from eli5.sklearn import PermutationImportance

• as_dataframe

  from eli5.formatters import as_dataframe

  Useful for programmatic access to explanation results, especially outside Jupyter notebooks.
• explain_weights
  wrong:

  from eli5.sklearn import explain_weights

  correct:

  import eli5
  eli5.show_weights(...)

  Old API: `explain_weights` and `explain_prediction` (and similar backend-specific functions) were replaced by generic `eli5.show_weights` and `eli5.show_prediction`.

Quickstart

This example demonstrates how to train a RandomForestRegressor on the California Housing dataset and then use `eli5.show_weights` to explain global feature importances and `eli5.show_prediction` to explain a specific prediction. Note the use of `feature_names` for clearer output.

import eli5
from sklearn.datasets import fetch_california_housing
from sklearn.ensemble import RandomForestRegressor
from sklearn.model_selection import train_test_split

# Load a modern dataset with feature names
housing = fetch_california_housing(as_frame=True)
X, y = housing.data, housing.target
X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=42)

# Train a simple model
model = RandomForestRegressor(random_state=42, n_estimators=10)
model.fit(X_train, y_train)

print("\n--- Feature Importances (Weights) ---")
# Explain feature importances (weights)
# In a Jupyter notebook, this would render as HTML.
# For console output, the string representation is printed.
print(eli5.show_weights(model, feature_names=X.columns.tolist()))

print("\n--- Prediction Explanation for First Test Sample ---")
# Explain a single prediction
print(eli5.show_prediction(model, X_test.iloc[0], feature_names=X.columns.tolist()))