Classical ML, validation, metrics, interpretation, debugging · insight-1166
Guard tree-model inputs against values outside the training range, because predictions go flat instead of failing
- Evidence type
- documented fact
- Difficulty
- medium
- Roles
- Data scientist, ML engineer
- Industries
- aviation, retail
Summary: situation, action, reason, boundary
Situation
A numeric feature at scoring time can exceed anything seen in training, such as a larger aircraft type or a new price tier.
Action
Store each numeric feature's training minimum and maximum with the model, flag scoring rows outside them, and send those to a fallback such as a linear baseline or a rule. If this is routine, fit a linear term first and boost the residual.
Reason
Tree predictions are piecewise constant, so a point beyond the last split lands in the edge leaf and inherits its value.
Boundary
Size, price, tenure and volume features that grow over time.
Symptom or risk
A tree or boosted model returns its outermost leaf's value, so the output stops responding to the feature yet looks plausible.
Exceptions
Features capped by design, such as a load factor that cannot exceed 1, need no range guard.
Trade-offs
A fallback adds a second model to maintain; residual boosting is only as good as the linear part.
Pitfalls
Checking each feature alone; a row can sit inside every marginal range and still be far from the training data.
Example
Hypothetical: a model trained on 120 to 189 seat aircraft scores a 300-seat widebody and returns what it learned for the 189-seat edge.
Lessons and concepts
Claims and their sources
Decision-tree predictions are piecewise constant approximations and trees are not good at extrapolation; a tree fitted on inputs 0 to 5 predicts the largest training target for an input of 10.