The Detective's Notebook

Opening Scene

A detective’s notebook doesn’t just rank which factors matter most overall, and it doesn’t only record the fine detail of one case. It also tracks patterns across a range: how does the likelihood of a particular outcome change as one specific factor moves from low to high, holding everything else roughly constant? That’s a genuinely useful middle ground between a single overall importance ranking and one case’s specific timeline.

That middle ground is exactly what a partial dependence plot shows.

In Plain English

A partial dependence plot (PDP) shows how a model’s predictions change, on average, as one feature varies across its range, while other features are held at their typical values. It reveals the shape of a feature’s influence — does risk rise steadily with age, spike after a threshold, or plateau — in a way that a single feature-importance number can’t capture on its own.

The Old Way

Before PDPs formalized this, people traced input-output relationships through cruder methods:

A scientist plotting one variable against an outcome while trying to hold everything else constant, the classic controlled-experiment approach, done manually and one variable at a time.
An economist charting how demand shifts as price changes, holding other market conditions fixed as a simplifying assumption.
A doctor noting how risk of a condition changes with age, based on aggregated case history rather than a systematic model-based method.

The underlying instinct — trace one variable’s effect while holding others fixed — long predates any formal machine learning technique.

What’s Changing (and Why AI Is the Reason)

PDPs can now be computed directly from any trained model, without needing a separate controlled experiment or new data collection, making this kind of analysis dramatically cheaper than its pre-model equivalents.
PDPs are increasingly paired with individual conditional expectation (ICE) plots, which show the same relationship for individual data points rather than just the average, catching cases where the average shape hides real variation underneath.
As models take on more features through the feature engineering techniques covered elsewhere in this content library, PDPs have become a standard way to sanity-check that a feature’s learned relationship actually makes real-world sense.

The Metaphor, Fully Extended

The Investigation	Partial Dependence Concept
Tracking how one factor’s influence changes across its range	Tracking how predictions change as one feature varies
Holding the rest of the case details roughly typical	Holding other features at their typical values
A pattern found in the notebook, not tied to one single case	The average relationship shown by a PDP
Checking individual case notes against the general pattern	Comparing ICE plots against the average PDP shape

For Beginners: What to Actually Do

Generate a PDP for a feature you already know is important, and check whether its shape matches your real-world intuition.
Learn to spot a PDP shape that doesn’t make domain sense — this is often a sign of a data or modeling problem worth investigating.
Pair PDPs with ICE plots when possible; a flat average PDP can sometimes hide sharply different behavior across subgroups.

For Practitioners and Leaders: The Deeper Layer

Use PDPs as a standard sanity check before deployment, especially for features with real-world policy implications, like age or income thresholds.
Watch for PDP shapes that suggest a model has learned an implausible or undesirable relationship, and investigate before it reaches production.
Recognize PDPs assume feature independence, which isn’t always realistic — a limitation worth naming explicitly to stakeholders.

Quick Recap

Partial dependence plots show how predictions change as one feature varies, holding others roughly constant.
They reveal the shape of a feature’s influence, which a single importance number can’t capture.
PDPs are computed directly from a trained model, without new data collection.
ICE plots complement PDPs by showing individual-level relationships the average might hide.

Where This Fits in the Series

Article 9 covered a middle-ground view between global and local explanations. Article 10 turns the investigation on itself, testing whether these explanations can be trusted under pressure.