Interactions plot parity with contribution plot (multiclass label, crossed sampling, jittering)

[guillaume-vignal]

Summary

interactions_plot and top_interactions_plot are not fully aligned with contribution_plot for multiclass behavior and interaction readability.

Problem

• Multiclass interaction values may come from SHAP in output-specific formats (list/4D), which can break plotting if not normalized.
• Interaction analysis should support class selection (label) like contribution_plot.
• Smart sampling for interactions should preserve the joint distribution of the 2 selected variables.
• Violin interaction rendering should include jittered point dispersion for readability while preserving true hover values.
• Postprocessed/display values should be consistently used in interaction axes and hover.

Expected Behavior

• interactions_plot and top_interactions_plot accept label and display class-specific interactions in classification.
• SHAP interaction outputs are robustly handled for multiclass/multi-output formats.
• Smart sampling for interactions is based on crossed variables.
• Violin mode shows jittered points with non-jittered values in hover.
• Postprocessed values are consistently reflected in interaction visual outputs.

Impact

This improves parity with contribution_plot, prevents multiclass edge-case failures, and makes interaction plots more interpretable and stable.