f_distance

Note Master 25-07-2026

How close the candidate is to the center of the stay. The score is 1 when the candidate is at the center and falls toward 0 as it gets farther away. How fast it falls depends on how spread out the stay's points are (\(\sigma_i\)) and on a setting \(\alpha\):

\[f_{\text{distance}}(E_k, C_i) = \exp\left(-\frac{d_{ki}^2}{2(\alpha \cdot \sigma_i)^2}\right)\]

\(d_{ki}\) is the distance from the stay's medoid to the nearest edge of candidate \(E_k\). The medoid is the GPS point closest to the average position of the stay. \(\sigma_i\) is the spread of the stay's points (sigma), with a minimum of 5 m. When the points are widely spread, the GPS position is less certain, so the score falls more slowly and farther places still get some credit.

In the models: \(\alpha\) was 1.0 (not tuned) until Alpha calibration set it to 2.4297. Before that, its weight changed sign between the 10 m and 15 m search distances (Model 1: base, Model 2: share-type and sigma). A squared version of this score also looked useful, but only because α was wrong (The f_distance² artifact).

Part of Framework overview.

Linked from Alpha calibration · Assignment framework · Framework overview · Model 1: base · Model 2: share-type and sigma · sigma