[ice-basic] Individual Conditional Expectation (ICE) Plot

[MarkusNeusinger]

Description

An ICE plot shows how the predicted outcome of a model changes as a single feature varies, with one line per observation. Unlike partial dependence plots (PDP) that show the average effect, ICE plots reveal heterogeneous effects and interactions by displaying individual prediction curves. This is an important complement to the existing pdp-basic spec.

Applications

• Detecting feature interactions in ensemble models
• Identifying subgroups with different feature responses
• Validating monotonicity assumptions in pricing models
• Exploring non-linear effects in clinical prediction models

Data

• feature_value (float) — grid of values for the feature of interest
• prediction (float) — predicted outcome for each observation at each grid point
• observation_id (int) — identifier for each observation
• Size: 50–200 observations × 50–100 grid points

Notes

• Individual lines should be semi-transparent to show density
• Optionally overlay the PDP (bold average line)
• Centered ICE variant (c-ICE) shifts all lines to start at zero

[claude]

✅ Spec ID: ice-basic

Rationale: ICE (Individual Conditional Expectation) plots are a well-established ML interpretability tool that complements the existing pdp-basic spec by showing per-observation prediction curves instead of the average. Valid and useful addition.

No duplicates found — pdp-basic is related but distinct (average vs. individual curves).

[github-actions]

✅ Specification Ready: ice-basic

PR: #5252

---

specification.md

ice-basic: Individual Conditional Expectation (ICE) Plot

Description

An Individual Conditional Expectation (ICE) plot visualizes how the predicted outcome of a machine learning model changes for each individual observation as a single feature varies across its range. Unlike partial dependence plots (PDP) that show the average marginal effect, ICE plots display one line per observation, revealing heterogeneous effects, feature interactions, and subgroup-specific behaviors that would be hidden by averaging. This makes ICE plots essential for detecting when a feature's effect varies across the population.

Applications

• Detecting feature interactions in ensemble models by identifying observations whose prediction curves diverge from the average trend
• Identifying subgroups with different feature responses in clinical prediction models, such as patients who respond differently to dosage changes
• Validating monotonicity assumptions in pricing or risk models by checking whether all individual curves move in the expected direction
• Exploring non-linear and heterogeneous effects in gradient boosting or random forest models during model audit and interpretability review

Data

• feature_value (float) - Grid of values for the feature of interest, spanning its observed range on the x-axis
• prediction (float) - Model-predicted outcome for each observation at each grid point on the y-axis
• observation_id (int) - Unique identifier for each observation, defining individual ICE lines
• Size: 50-200 observations with 50-100 grid points per observation
• Example: ICE curves from a GradientBoostingRegressor showing how house price predictions change as square footage varies for individual houses

Notes

• Individual lines should use low alpha (semi-transparent) to show density and overlap patterns
• Overlay the partial dependence (PDP) line as a bold, opaque curve to show the average effect
• Consider a centered ICE (c-ICE) variant that shifts all lines to start at zero for easier comparison of effect shapes
• A rug plot along the x-axis can indicate the distribution of observed feature values
• Color-coding lines by a second feature can help reveal interaction effects
• Complements the existing pdp-basic spec by showing individual-level detail behind the average

---

Next: Add approved label to merge this specification to main.

---

🤖 spec-create workflow

[github-actions]

🚀 Specification Merged!

Specification: ice-basic
PR: #5252 (merged)

The specification is now in main and synced to PostgreSQL.

---

Next: Add generate:{library} labels to start generating implementations.

Available libraries: matplotlib, seaborn, plotly, bokeh, altair, plotnine, pygal, highcharts, letsplot

Or use workflow dispatch for bulk operations.

---

🤖 spec-create workflow