Enhanced Explainability: Display Prediction Correctness in Contribution Plot

[guillaume-vignal]

Description:

We propose an enhancement to the contribution_plot in Shapash to improve model explainability by visually representing the correctness of each prediction. This will provide a more intuitive understanding of model behavior and facilitate error analysis.

Feature Overview

• Visual differentiation based on true class:

  • Use different marker shapes depending on the true target class (e.g., circle for class 0, square for class 1, etc.).

• Flexible color mapping:

  • Allow the user to switch the color encoding between:

    • Model predictions
    • True target values
    • Prediction errors (e.g., correct vs incorrect)

  • The available modes should adapt to the type of model (classification vs regression).

• Integration context:

  • This feature should be available:

    • In the notebook visualizations via additional parameters to the plotting function.
    • In the webapp via an interactive menu or toggle options.

Expected Benefits

• Easier identification of where the model is performing well or poorly.
• Improved user control over the visualization semantics.
• More detailed diagnostic capability in both development and presentation environments.