Native XAI Support with SHAP Values and LIME Explanations in PMML

[olowonosmall]

Description:
Integrate explainable AI techniques directly into PMML exports by embedding SHAP
values, LIME explanations, and other interpretability artifacts. This would make
models self-documenting and enable downstream systems to provide explanations
without requiring the original Python environment.

Technical Scope:

• SHAP value computation and embedding in PMML for each feature
• LIME local explanation generation and export
• Feature importance scores with confidence intervals
• Partial dependence plots (PDP) and Individual Conditional Expectation (ICE) data
• Counterfactual explanation generation and export
• Anchors and rule-based explanations for complex models
• Attention weights for neural network models
• Global and local explanation metadata in PMML MiningSchema

Implementation Challenges:

• Extending PMML schema to accommodate explanation data
• Balancing model size with explanation detail
• Ensuring explanation accuracy across different PMML consumers
• Computing explanations efficiently for large models

Use Cases:

• Regulated industries requiring model explainability (finance, healthcare)
• Customer-facing applications needing prediction justifications
• Model debugging and validation workflows
• Compliance with AI transparency regulations (EU AI Act, GDPR)

[vruusmann]

So we have an AI requesting that I write an XAI interface for it?

The SkLearn2PMML package already provides the "PMML customization" mechanism, which allows you to add/modify/remove parts of the PMML document while it is being generated. Thus, your use case (attach field metadata such as LIME or SHAP values) would be addressed by adding a custom PMML fragment into the document.

Here are two examples about attaching model quality data:
https://github.com/jpmml/jpmml-sklearn/blob/1.9.18/pmml-sklearn/src/test/resources/main.py#L183
https://github.com/jpmml/jpmml-sklearn/blob/1.9.18/pmml-sklearn/src/test/resources/main.py#L850

See also:
#410

Admittedly, the existing API is a bit too difficult for regular users. The PMMLPipeline class could provide some helper methods for generating and attaching model quality information more "atomically".

Also, the PMMLPipeline class could provide some more extra methods for decorating field declarations with metainfo. So the user gets to compute her favourite importance metrics/scores using her favourite Python tools/libraries, and then simply embed these results into the PMML document.