Model 1: base
The first and simplest model: the four scores, each with one weight shared by all place types.
\[V_k = w_1 f_\text{distribution} + w_2 f_\text{distance} + w_3 f_\text{temporal} + w_4 f_\text{directional}\]
Tested with two search distances, 10 m and 15 m:
| R=10 m | R=15 m | |
|---|---|---|
| LL(C) / LL(b) | -0.9362 / -0.9745 | -0.9985 / -1.0360 |
| ρ² vs LL(C) | -0.041 | -0.038 |
| Accuracy | 0.518 | 0.509 |
| Building / street / open space | 0.576 / 0.378 / 0.492 | 0.589 / 0.346 / 0.445 |
| CV accuracy (grouped) | 0.514 ± 0.089 | 0.512 ± 0.098 |
| Condition number | 36 | 26 |
| Term (R=10 m) | w | t | VIF |
|---|---|---|---|
| f_distribution | +3.331 | 5.34 | 7.14 |
| f_distance | -0.467 | -1.14 | 3.73 |
| f_temporal | +0.179 | 0.29 | 7.49 |
| f_directional | -0.080 | -0.41 | 2.78 |
- Worse than a simple baseline. ρ² is negative, so this model does worse than just knowing how common each type is. No value of α fixes this (Alpha calibration).
- Only f_distribution is significant at both distances. f_temporal moves almost the same way as f_distribution (r ≈ 0.88), so it was removed.
- f_distance has a negative weight at 10 m (-0.47) and a positive one at 15 m (+0.95). This had to wait until α was tuned.
- The condition number is low, so the fit itself is stable. The problem is which inputs the model has, not the fitting.
Next: Model 2: share-type and sigma. Part of Calibration.