58 Prediction and interpretation are two uses of a model

What people usually ask

Which inputs matter most?

What comes back

A ranked bar chart of importances with no word on which inputs move together.

Ask this instead

I have crystallizer_runs.csv: 38 runs, inputs cooling_rate,
seed_mass, agitation_rpm, hold_time, final_temp, target d50_um.
First ask me which of two goals this serves and wait: predicting
d50 at new settings, or understanding which settings to change.
Write the answer at the top of the notebook.
Then fit the pipeline under 5 fold cross validation. On the held
out rows of each fold, compute permutation importance with 10
repeats per input. Report each input's mean drop in R2 with its
standard deviation across folds and repeats, as a table sorted
by mean.
Beside it, print the correlation matrix of the five inputs and
mark every pair above 0.5.
Under both, state that importance here means association within
these 38 runs: an input can rank high because it tracks another
input or the run order, not because it sets d50.
For the top two inputs, write the one experiment that would
settle whether each causes the change: what to hold, what to
vary, how many runs. Do not run it.

Attach: crystallizer_runs.csv.

Why it works: The ranking is tied to the goal it serves, carries its spread and its correlations, and the question of cause goes to an experiment that is written down rather than assumed.