PDP report

1 · Overview — where to look

An additive surrogate read off the global curves reproduces 54.2% of the model's predicted variance; adding the regional plots kept by the decision sequence, 82.5%.

Each point is a feature: importance (x) against heterogeneity (y). Bottom-left is ignorable; bottom-right is important and fully described by its mean effect; the top-right corner — important and heterogeneous — is where the mean hides something. An arrow marks each split the decision sequence accepted: from the feature's global point to its weighted-mean point across the subregions.

The plotted features carry 87% of the total importance mass (target 80%, ceiling top_k = 5).

The decision sequence. Starting from the global curves, each round applies the split with the largest explained-variance gain, measured on top of the splits above it, and stops when no remaining split adds at least 1.0%. A real split (its heterogeneity does drop) can still add nothing — or even hurt, by double-counting — when its variance is already explained by an earlier split.

Bar view — importance and heterogeneity

2 · Regional analysis — the final CALM

The selected snapshot: global effects everywhere except the accepted splits. Features in descending importance — a split feature enters as one group at the instance-weighted mean of its subregions. The split features' global counterparts are in the baseline section at the end.

2.1 · frequency

Split on chord-length, suction-side-displacement-thickness into 4 regions — worth +28.3% of explained variance on top of the splits above it; importance and heterogeneity here are the instance-weighted means over the subregions. The global counterpart is in the baseline.

Feature 0 - Full partition tree:
🌳 Full Tree Structure:
───────────────────────
frequency 🔹 [id: 0 | heter: 5.18 | inst: 1202 | w: 1.00]
    suction-side-displacement-thickness < 0.00 🔹 [id: 1 | heter: 4.22 | inst: 400 | w: 0.33]
        chord-length < 0.05 🔹 [id: 2 | heter: 2.19 | inst: 141 | w: 0.12]
        chord-length ≥ 0.05 🔹 [id: 3 | heter: 1.93 | inst: 259 | w: 0.22]
    suction-side-displacement-thickness ≥ 0.00 🔹 [id: 4 | heter: 4.18 | inst: 802 | w: 0.67]
        chord-length < 0.04 🔹 [id: 5 | heter: 2.27 | inst: 145 | w: 0.12]
        chord-length ≥ 0.04 🔹 [id: 6 | heter: 3.09 | inst: 657 | w: 0.55]
--------------------------------------------------
Feature 0 - Statistics per tree level:
🌳 Tree Summary:
─────────────────
Level 0🔹heter: 5.18
    Level 1🔹heter: 4.19 | 🔻0.99 (19.04%)
        Level 2🔹heter: 2.64 | 🔻1.56 (37.14%)

2.2 · suction-side-displacement-thickness

Global effect

Regional effects

A split on frequency into 3 regions was found (heterogeneity 4.322 → 2.439), but the decision sequence skips it: it adds no explained variance beyond the splits kept there — the same variance is already read elsewhere. The regional plots are omitted; reproduce them with find_regions.

2.3 · chord-length

Global effect

Regional effects

A split on frequency into 3 regions was found (heterogeneity 2.345 → 1.418), but the decision sequence skips it: it adds no explained variance beyond the splits kept there — the same variance is already read elsewhere. The regional plots are omitted; reproduce them with find_regions.

3 · Global baseline — without regions

What you would believe about the split features without the regional analysis: their global mean effects, with the heterogeneity the accepted splits just explained still hiding inside the band. Compare with their subregions in the regional analysis above.

frequency