Classical ML, validation, metrics, interpretation, debugging · insight-0185
Plot ICE curves with the PDP so averages do not hide opposite effects
- Evidence type
- documented fact
- Difficulty
- medium
- Roles
- Data scientist, Data analyst
- Industries
- aviation, retail
Summary: situation, action, reason, boundary
Situation
A partial dependence plot shows a flat or mild effect for a feature that domain experts expect to matter.
Action
Overlay individual conditional expectation curves (kind='both'), center them when many lines overlap, and colour them by a suspected moderator such as market.
Reason
A PDP averages predictions over the data, so heterogeneous or interacting effects can average out; ICE shows one curve per instance.
Boundary
Tree ensembles and other flexible models where interactions are likely.
Symptom or risk
Opposite effects in subgroups cancel in the average; the PDP suggests no effect while individual predictions move strongly.
Exceptions
For additive models without interactions, ICE curves are parallel and the PDP suffices.
Trade-offs
ICE plots get cluttered on large samples; subsample and use centered curves.
Pitfalls
Forgetting that both PDP and ICE assume independence from other features and may show impossible combinations.
Example
Synthetic airline data: true price elasticities differ by route (about -0.6 on JFK to -1.35 on AYT), so a pooled PDP averages them into one curve while ICE curves coloured by route can show the spread.
Lessons and concepts
Claims and their sources
ICE plots show heterogeneous effects hidden by PDP averages; scikit-learn supports kind='both' and centered ICE.