A four-state partial dependence plot with the ICE bundle it is the average of — one faint line per instance under the bold mean, a centring toggle that re-bases every curve on the first grid point, curves that move against the average dashed and counted, and a decile rug showing where the training data actually lives.
Preview in your theme
Loading preview…
This component is part of Pro
The install command, prompt and concept map are included with a Pro license — one purchase unlocks every Pro component, forever.
status="ready" · a churn model re-scored across thirteen tenure values for a simulated book of 600 customers. Each hairline is one customer with every other feature held at their own values; the bold line is the average of all 600, which is all a partial dependence plot is. Five of the twenty-eight drawn curves rise while the average falls — the two-year cohort, whose risk climbs toward the renewal price — so they are dashed and counted rather than left to disappear into the bundle. Press Centred to re-base every curve on its value at month 1: the levels collapse and the disagreement is all that is left. Point at the plot, or Tab to it and use the arrow keys.
Churn model · partial dependence on tenure
Tenure 1–60 mo · average 0.38 → 0.08 (-0.30) · 5 of 28 curves (18%) move the other way
Tenure 24 mo · average 0.21 · middle 80% 0.06 to 0.38 · 8.7% of rows near here
partial dependence · the averaged prediction
ICE · one curve per instance (28)
moves the other way · 5 (18%)
600 training rows · tall ticks are deciles
Each faint line is one instance re-scored across the Tenure grid with every other feature left at that row's own values; the bold line is their average, which is all a partial dependence plot ever is — PD(v) = mean of the curves at v. On this axis the bundle's own mean sits up to 0.01 from the plotted average — usually just the price of drawing a subsample, and worth checking when it is not small. Press Centred to re-base every curve on its own value at 1 mo — levels collapse and only the disagreement about the effect is left. 5 of 28 curves end higher than they started while the average moves -0.30 — the average is hiding them, which is the entire reason to draw the bundle. The rug is where the Tenure column actually lives: 300 ticks spaced evenly by rank through the 545 rows in range, a taller tick at each decile, so a tenth of the rows sits between two tall ticks. Sampling by rank keeps the empirical quantiles, so the thinned rug is as dense as the full column would have drawn it. 55 of 600 rows lie outside the plotted sweep and are counted here rather than pushed onto the edge. Point at the plot, or Tab to it and use the arrow keys, to read any grid point.
Tenure 1–60 mo · average 0.38 → 0.08 (-0.30) · 5 of 28 curves (18%) move the other way. Up to 14 grid points, evenly spaced by rank.
Tenure
Partial dependence · P(churn within 12 months)
Curves 10th percentile
Curves median
Curves 90th percentile
Training rows near this point
1 mo
0.38
0.04
0.37
0.68
53 of 600
3 mo
0.36
0.05
0.36
0.65
67 of 600
6 mo
0.34
0.05
0.33
0.61
44 of 600
9 mo
0.32
0.05
0.31
0.57
28 of 600
12 mo
0.30
0.06
0.28
0.52
51 of 600
18 mo
0.25
0.06
0.24
0.44
52 of 600
24 mo
0.21
0.06
0.21
0.38
52 of 600
30 mo
0.18
0.06
0.17
0.32
36 of 600
36 mo
0.15
0.05
0.15
0.25
41 of 600
42 mo
0.12
0.05
0.11
0.22
36 of 600
48 mo
0.10
0.04
0.08
0.21
29 of 600
54 mo
0.09
0.04
0.07
0.21
32 of 600
60 mo
0.08
0.03
0.06
0.22
24 of 600
status="empty" · a valid contract with no grid at all, worded so it cannot be mistaken for a failed fetch — the feature is known, the sweep is not.
Churn model · partial dependence on tenure
Average model response to Tenure, drawn over one curve per instance
No grid to sweep
A partial dependence plot needs the model re-scored at a sweep of Tenure values — at least two of them, with every other feature held at each row's own values. A single importance number cannot be unrolled back into the shape of the response.
status="loading" · a skeleton holding the plot box: a fan of individual curves under a bolder average, with the rug lane below, aria-hidden, honouring prefers-reduced-motion, and no timer involved.
Churn model · partial dependence on tenure
Average model response to Tenure, drawn over one curve per instance
Loading the partial dependence plot
status="error" · the fetch failed. Try again only appears when onRetry is passed, and clicking it flips the card to ready.
Churn model · partial dependence on tenure
Average model response to Tenure, drawn over one curve per instance
Zyeon UI