Skip to content

compare and plot_triage

Description

Part of the interactive API guide: the cross engine views. compare overlays methods on one feature, plot_triage draws the importance versus heterogeneity plane, FeatureEffect runs a pool of engines behind one handle.

Reading time

Approx. 4' to read.

compare: a second opinion

effector.compare(pdp, ale, feature="hr")

PDP vs ALE on hr

One axis, the mean effect of every engine you pass (always centered, so the curves are comparable). Two very different estimators agreeing this closely is the cheapest robustness check there is; where they diverge, the divergence itself is the finding (usually correlated features, which PDP extrapolates over and ALE does not).

It also compares models over the same columns: pass two fitted engines of the same method built on different models, with labels= to name them.

plot_triage: the plane, by hand

effector.plot_triage(pdp)                      # every feature, two axes
effector.plot_triage(pdp, partitions=parts)    # plus the before/after arrows

Triage with candidate arrows

Every feature sits at (importance, heter_score); the hairline marks the median heterogeneity. With partitions= (what find_regions returned), each found split draws one arrow per subregion, from the feature's global point to that region's own scores.

⚠️ Unlike the report's triage plane, which draws only the splits the decision sequence accepted, this one draws every partition you hand it: workingday's arrows are here, even though select_regions would reject its split as redundant. Candidates on this plot, verdicts on the report's.

FeatureEffect: a pool behind one handle

fe = effector.FeatureEffect(X_test, predict)   # pdp + ale + rhale + shapdp
fe.plot(0)                                     # all four, overlaid
fe.eval(0, xs)                                 # {"PDP": ..., "ALE": ...}
fe.plot(0, methods=["pdp", "shapdp"])          # a subset

The constructor mirrors the engines (schema, nof_instances, random_state, and model_jac to unlock rhale). DerPDP is excluded on purpose: its derivative scale curves cannot share an axis with the others. Per method arguments pass through method_kwargs.

Where to next