Skip to main content
An FPDE explanation is a target-versus-rival feature attribution. Read every value in that local contrast.

​
Attribution signs

Positive attribution values support the target class. Negative attribution values support the rival class. Values near zero contribute little to that local contrast.
The sign is relative to the selected target class and rival class. A feature can change sign under a different rival class.

​
Evidence and exactness

details["evidence"] is the scalar contrast decomposed by the attribution vector. For an exact decomposition, the attribution sum should be close to the evidence value.
Small residuals are expected because of floating-point arithmetic. Large residuals suggest a numerical or input-shape problem.

​
Target and rival labels

With FPDEEngine, the target class comes from the highest predict_proba value. The rival class comes from the second-highest predict_proba value. Use these labels when you explain results to readers. Avoid saying that an attribution explains every class at once.

​
Raw and normalized values

Use raw attributions when you report evidence-scale values. Use normalized attributions only for display or comparison.
If you call lower-level functions and receive an FPDEExplanation, use normalized_attributions for display-scale summaries. Keep attributions for the raw decomposition.

​
What not to claim

  • Do not describe FPDE as a causal explanation method.
  • Do not claim support for prototype types beyond class-mean prototypes in fpde 0.1.0.
  • Do not compare attribution signs across samples without checking each sample’s target and rival class.