DataTerminology.comNEW
Also known as: ICE Plots
Curves showing how a model’s prediction for one instance changes as a single feature varies, holding other features fixed.
DESCRIPTION
Individual Conditional Expectation (ICE) plots visualize instance-level prediction changes over a grid of feature values. They reveal heterogeneity masked in Partial Dependence Plots by presenting one curve per instance. ICE helps diagnose interactions, nonlinearity, and regions of instability. They support explainability and trustworthiness by making local effects explicit. Care must be taken with correlated features and extrapolation beyond observed ranges. Aggregating ICE yields PDP, while centered ICE improves interpretability.
ALTERNATIVE DEFINITIONS
Goldstein et al.
Individual conditional expectation (ICE) plots visualize the dependence of the prediction on a feature for individual instances.
Introduces ICE as instance-level visualization contrasting PDP.
Interpretable Machine Learning
ICE plots show how the prediction for a single instance changes when a feature is varied, keeping all other features at their observed values.
Emphasizes ceteris paribus condition for local analysis.
scikit-learn
Partial dependence plots can be complemented by individual conditional expectation curves to show heterogeneity of effects across samples.
Positions ICE alongside PDP within sklearn tooling.
RELATED TERMS
Accumulated Local Effects SHAP Values LIME Feature Effects Ceteris Paribus Profiles Model Explainability Global Surrogates Monotonic Constraints Sensitivity Analysis
COMPETING TERMS
RELATIONS
- Broader: Expectation
KNOWLEDGE AREAS
Data Science AI MLOps Testing & Validation
TARGET AUDIENCE
Data Scientist AI Engineer MLOps Engineer Chief AI Officer
EXTERNAL SOURCES
- Wikipedia - Partial dependence plot
- Goldstein et al. - Original ICE plots paper
- Christoph Molnar - Book chapter on ICE plots
- scikit-learn - PDP and ICE in sklearn
TOP QUESTIONS
How do ICE plots differ from Partial Dependence Plots?
ICE shows one curve per instance to reveal heterogeneous effects; PDP averages those curves, potentially masking interactions and variability.
Category: Distinction
What is a common mistake when using ICE plots?
Interpreting extrapolated regions where feature values are unrealistic for the instance. Restrict ranges to plausible values.
Category: Mistakes
How can ICE plots help with model trust?
They expose local behavior, highlighting where predictions are sensitive or unstable, enabling corrective actions like constraints or feature engineering.
Category: Relevancy
What is a hot topic around ICE plots?
Combining ICE with counterfactual explanations and robust methods under correlated features.
Category: Topic
How to interpret centered ICE plots?
Centering subtracts baseline prediction to focus on changes. It eases comparison across instances.
Category: HowTo
What risks exist in using ICE plots for correlated features?
Varying one feature while fixing others may create implausible combinations, misleading interpretation. Consider conditional ICE or ALE.
Category: Risks
Where should beginners start with ICE plots?
Read the original paper and Molnar’s chapter, then implement with sklearn’s inspection tools or custom plotting.
Category: HowTo
Do ICE plots scale to high-dimensional models?
Yes, but consider focusing on top features by importance and sampling instances to maintain readability.
Category: Management
What is the biggest misunderstanding about ICE plots?
Assuming monotonic curves imply causal effects. ICE is descriptive of model behavior, not causality.
Category: Misunderstanding