TreeExplainer interventional explanations become slow after the background grows from 500 to 50,000 rows. What should…
Classic MLLevel: IntermediateSHAP · TreeExplainer
TreeExplainer interventional explanations become slow after the background grows from 500 to 50,000 rows. What should be changed first?
Hints
- The background is an integration reference, not a mandatory training replay.
- Balance representativeness against evaluation cost.
Answer
Benchmark a representative smaller background rather than treating every training row as mandatory. Runtime scales with background size; preserve relevant population structure and measure explanation stability across samples.
Source and licence: SHAP · TreeExplainer · MIT · answer: Quantoros editorial