PDPbox Documentation

Individual Conditional Expectation (ICE) plots can suffer from two main visual issues:

1. Difficulty comparing curves: Because curves start at different $\hat{f}(x)$ values, it can be hard to see if they differ in shape. Use the center parameter in pdp_plot or pdp_interact_plot to align them.
2. Overcrowding: Drawing too many ICE curves makes the plot unreadable. Use the frac_to_plot or cluster parameters in pdp_plot or pdp_interact_plot to limit the number of curves displayed.

Omitting the feature distribution in Partial Dependence visualizations can be misleading, as it may lead to over-interpreting the plot in regions where there are almost no actual feature values.

To include the feature distribution in your plots, use the plot_pts_dist parameter in the pdp_plot method.

PDPbox follows a layered architecture designed to separate the user interface from the underlying plotting logic and visual styling. The architecture consists of three main layers:

1. Plot API: The primary entry point for users to interact with the library.
2. Plot Engine: A layer that contains the specific logic for generating different types of plots.
3. Plot Style: A layer dedicated to managing the visual appearance and styling of the plots.

Users interact with the Plot API, which delegates the heavy lifting to the appropriate Plot Engine, which in turn uses a Plot Style to ensure consistent and correct visual representation.

The project uses two distinct tools for coverage management:

1. Coverage.py: Controlled by .coveragerc. This file defines how coverage data is collected during execution.
2. Codecov: Controlled by codecov.yml. This file defines how the collected data is interpreted and presented in the Codecov dashboard.

Partial Dependence Plots (PDP) assume that the features being analyzed ($X_C$) are uncorrelated with the remaining features ($X_S$). If features are highly correlated (e.g., height and weight), the averaging process in PDP may include unrealistic data points, making the results untrustworthy.

To mitigate this, use the data_transformer parameter in the following methods to transform or filter the data used for calculation:

• pdp_isolate
• pdp_interact

PDPbox is designed to visualize how specific features influence model predictions using Partial Dependence Plots (PDP).

Scikit-learn Models

If you are using standard scikit-learn models, PDPbox automatically detects the prediction interface. It typically looks for:

• model.predict
• model.predict_proba

Custom Models

For models that do not follow the standard scikit-learn interface, you can provide a custom prediction function using the pred_func parameter available in PDPbox methods. This allows you to use PDPbox with any model by explicitly defining how predictions should be generated.

The pdpbox.info_plots module provides two main classes for visualizing interaction information: InteractTargetPlot and InteractPredictPlot. Both classes inherit from the internal _InteractInfoPlot base class and share a common .plot() method.

• Use InteractTargetPlot to visualize interaction information related to the target variable.
• Use InteractPredictPlot to visualize interaction information related to the model predictions.

# Example usage pattern (exact parameters depend on specific class implementation)
from pdpbox.info_plots import InteractTargetPlot, InteractPredictPlot

# For target-based interaction plots
plot_target = InteractTargetPlot(target_data, feature_data, ...)
plot_target.plot()

# For prediction-based interaction plots
plot_predict = InteractPredictPlot(model, feature_data, ...)
plot_predict.plot()

To install the latest development version from the GitHub repository, clone the repository and run the setup script.

$ git clone https://github.com/SauceCat/PDPbox.git
$ cd PDPbox
$ python setup.py install

You can analyze the Python code and generate UML diagrams using pyreverse (part of the PyLint tool suite). This command generates diagrams in PlantUML format and saves them to the specified directory.

pyreverse -o plantuml -d assets/diagrams pdpbox

After recording parameters in parameters.json and inserting placeholders in your Python files, run the generate_docstring.py script to automate the process.

Steps:

1. Navigate to the helper directory.
2. Execute the script providing the parameter file and the target Python files.
3. Review the generated *_updated.py files using a diff tool.
4. Replace the original files with the updated versions if the changes are correct.

cd assets/docs_helper
python generate_docstring.py --param_file parameters.json --py_files <py_file1> <py_file2> ...