Extend Local Plot to Support Subset of Selected Instances

[guillaume-vignal]

Description

Currently, the Shapash Local Plot displays the Shapley contributions for a single selected individual. This is extremely useful for local interpretability, but it becomes limiting when users interactively select a region of points (subset of individuals) from projection or other scatter plots.

The application already allows users to:

• Select a region (lasso or rectangular selection) of points in the Contributions Projection Plot.
• Select multiple points in other graphical components.

However, the Local Plot still only reflects the explanation of one individual at a time.

Feature Proposal

Extend the Local Plot functionality to support the visualization of Shapley contributions for a subset of selected individuals, instead of a single instance only.

This new feature would enable users to dynamically switch between:

• Single instance explanation (current behavior)
• Subset explanation (new behavior)

Expected Behavior

When a region or multiple points are selected:

1. The Local Plot should update to display aggregated Shapley contributions for the selected subset.

2. Possible aggregation strategies could include:

   • Mean absolute Shapley values
   • Median contributions
   • Distribution (boxplot )

3. A toggle or automatic mode could be used:

   • Single selection leads to the current local plot
   • Multiple selection leads to an aggregated local plot

Use Cases

• Analyze why a specific cluster behaves abnormally.
• Understand the dominant features driving predictions for a group of similar instances.
• Compare different regions of the projection space.

Implementation Notes

• Detect multi-point selections from interactive plots.
• When multiple points are selected, compute aggregated Shapley values over the selected indices.
• Adapt the Local Plot component to display group-level explanations.
• Provide clear visual distinction between single and group explanations.

UI / UX Suggestions

• Display a label such as:
  "Explanation for 24 selected instances"
• Add an uncertainty/standard deviation band to the plot for each feature

Expected Benefits

• Enhanced interpretability at cluster level.
• Better understanding of model behavior on specific regions of the data.
• Strengthened analytical workflow between global and local explanations.

---

This enhancement would significantly improve exploratory analysis by enabling users to interpret not only individual predictions but also collective behaviors across selected groups of instances.