Overfitting signature

Note Master 25-07-2026

Again and again, a more complex model fit the training data better without doing better on unseen data. That is the sign of overfitting: the model learns details of this sample that do not carry over. Three cases:

1. Adding an ASC term (a fixed extra value per type): Model 5: alternative-specific constants (rejected)
2. Giving f_directional a separate weight per type: f_directional: shared vs split
3. The squared distance term, before α was tuned: The f_distance² artifact

The tree models show the same thing on a bigger scale. Both beat the simple model on the training data and lose to it on unseen data (ML ceiling check, row level).

Why this matters for MILES: the survey's mix of 59% building, 20% street and 21% open space was set by the survey plan (Ground truth dataset). The share_type weights learned patterns of this sample. MILES may have a different mix, so it is not yet known how well the model will work there (Limitations).

Answers Research questions RQ3. Part of Complexity and transfer.

Linked from Complexity and transfer · Future work · Ground truth dataset · Limitations · ML ceiling check, row level · Model 5: alternative-specific constants (rejected) · Research questions · The f_distance² artifact · f_directional: shared vs split