Calculator Inputs
Formula Used
The calculator evaluates one model score for each observation and grid value. Other observation characteristics remain fixed through modifiers and offsets. The selected link converts each score into a prediction.
Identity links return scores without transformation. Logistic links convert scores into probabilities. Exponential links produce positive model responses.
How to Use This Calculator
Enter the tested feature range and desired grid size. Choose a model response and set its coefficients. Add observation modifiers that represent changing interaction conditions.
Enter optional offsets for different baseline predictions. Select centering, averaging, derivatives, scaling, clipping, or markers. Press the calculation button to display results above the form.
Compare line shapes, crossings, slopes, and endpoint spread. Review the average curve beside individual responses. Adjust one parameter gradually when exploring model behavior.
Understanding Individual Conditional Expectation
Why ICE Curves Matter
Individual Conditional Expectation plots clearly reveal how one feature changes each prediction. Every line represents one separate observation while other feature values remain fixed. This structure exposes important response differences hidden by a single average curve.
A partial dependence curve summarizes all observations with one combined mean response. That average may conceal subtle subgroups, interactions, thresholds, and opposing model behavior. Individual ICE lines preserve those patterns and support richer model interpretation.
Configurable Model Behavior
This calculator generates realistic synthetic predictions from configurable mathematical model components. Users control adjustable linear, quadratic, interaction, offset, and link parameters. These options reproduce many common shapes found in practical machine learning.
The feature grid defines every tested value along the horizontal axis. More grid points create smoother lines but require additional calculations. Wider ranges can reveal saturation, instability, or unrealistic extrapolation behavior.
Observation modifiers create different interaction strengths across simulated records. Observation offsets shift baseline predictions without changing basic response shapes. Together, these inputs demonstrate population heterogeneity with transparent assumptions.
Centering and Average Effects
Centered ICE subtracts each line's starting prediction from later predictions. This transformation emphasizes shape differences instead of baseline prediction differences. It often makes interactions and subgroup behavior easier to recognize.
The average curve provides a partial dependence style summary. Compare that curve with individual lines before making broad conclusions. Large separation indicates that the average may oversimplify important behavior.
Derivatives and Link Functions
Derivative values estimate local sensitivity across neighboring feature grid points. Steep derivatives indicate regions where predictions change especially quickly. Sign reversals can identify turning points created by quadratic effects.
Logistic links convert raw model scores into probabilities between zero and one. Exponential links create positive responses useful for count style outcomes. Identity links preserve unrestricted numerical predictions for regression examples.
Scaling and Output Controls
Scaling changes how feature values enter the selected mathematical formula. Standardization uses a supplied mean and standard deviation before prediction. Incorrect scaling assumptions can produce misleading curve shapes and magnitudes.
Clipping limits displayed predictions when extreme formulas create unrealistic values. Use clipping carefully because it can hide genuine model instability. Review unclipped outputs whenever unexpected plateaus appear across many lines.
Responsible Interpretation
ICE interpretation should consider data support around every tested feature value. Curves outside observed ranges may describe unsupported counterfactual combinations. Domain knowledge remains essential when judging whether patterns are credible.
Strong line crossings usually suggest interactions with other observation characteristics. Parallel lines usually indicate similar feature effects across sampled records. Mixed patterns may justify segmentation or additional interaction analysis.
Frequently Asked Questions
1. What does each ICE line represent?
Each line represents one observation across changing feature values. Other modeled characteristics remain fixed for that observation.
2. How does ICE differ from partial dependence?
ICE displays individual prediction paths. Partial dependence averages those paths into one summary curve. The average can hide important variation.
3. Why should I center ICE curves?
Centering removes different starting predictions. It highlights shape changes, interactions, and response differences more clearly.
4. What do crossing ICE lines suggest?
Crossings often suggest interactions or changing subgroup effects. They can also reveal nonlinear behavior across observations.
5. When should logistic response be selected?
Select logistic response when predictions represent probabilities. Outputs remain between zero and one.
6. What does the derivative chart show?
The derivative estimates local prediction sensitivity. Larger absolute values indicate faster changes across nearby feature values.
7. Why use observation modifiers?
Modifiers create different interaction strengths. They help simulate heterogeneous responses among observations.
8. Can this replace production model explanations?
No. This calculator demonstrates ICE mechanics with a configurable synthetic model. Production explanations require actual model predictions.
9. How should extreme curves be interpreted?
Check coefficients, scaling, clipping, and feature ranges first. Extreme responses may indicate extrapolation or unstable assumptions. Always compare explanations with data support and domain knowledge.