Partial Dependence Plot (PDP)

A partial dependence plot (PDP) shows the average predicted outcome of a fitted machine learning model as one or two features vary, with the effects of all other features averaged out.

부분의존도 그래프(PDP)는 다른 모든 특성의 효과를 평균한 상태에서 한두 개 특성이 변할 때 학습된 머신러닝 모형의 평균 예측값이 어떻게 변하는지를 보여준다.

How do I see how a feature affects the predictions of a black-box machine learning model?

블랙박스 머신러닝 모형에서 특정 변수가 예측에 어떤 영향을 주는지 어떻게 보나요?

A partial dependence plot shows how a model's average prediction changes as one or two features vary, which makes the learned relationship visible for black-box models. It fits global interpretation of tree ensembles or neural networks, and individual conditional expectation curves reveal effects that differ across observations.

When to use Partial Dependence Plot (PDP)

• A trained model such as a random forest, gradient boosting or neural network
• Explaining the shape of a feature's effect on predictions
• Examining two-way feature interactions

When to use something else

• Strongly correlated features, where averaging creates unrealistic combinations; use accumulated local effects plots
• Explaining a single prediction; use SHAP values or LIME

Assumptions and requirements

• A fitted predictive model
• Features not strongly correlated with the others
• Average effect may hide heterogeneous effects

What the analysis reports

• One-way partial dependence curves
• Two-way partial dependence heat maps
• Individual conditional expectation (ICE) curves

How to report it

The partial dependence plot showed that predicted [outcome] increased with [feature] up to approximately [value] and leveled off thereafter.

Related methods

• SHAP Values
• Accumulated Local Effects
• Individual Conditional Expectation

Run Partial Dependence Plot (PDP) on Quorant