sigma
How spread out the GPS points of one stay are. A large σ means the points are scattered, so the GPS position was uncertain.
\[\sigma_i = \sqrt{\mathrm{var}(x) + \mathrm{var}(y)}\]
\(\mathrm{var}(x)\) and \(\mathrm{var}(y)\) are the variances of the east-west and north-south positions.
σ is used in two places:
- Inside f_distance it sets how fast the distance score falls, with a minimum of 5 m.
- As its own model input. Very large values are capped at 179.14 m (the 99th percentile). Then the value is divided by
sigma_scale= 36.708, the standard deviation of the capped values. This scale is computed once and reused unchanged in every model:
\[\sigma_k = \begin{cases} \min(\sigma_i,\ 179.14) \,/\, 36.708 & \text{street or open-space candidate} \\ 0 & \text{building candidate} \end{cases}\]
σ gets one weight shared by all types, with building as the reference (0 for building rows). σ has the same value for all candidates of a stay. Giving it a separate weight per type would make the model unable to tell the weights apart, so that option is not available.
In the models: σ has a clear positive effect in every model (Model 2: share-type and sigma, Final model). Tree models rank it much higher than its small weight suggests: Sigma in tree importance.
Part of Assignment framework.