Directional scope ablation (3a vs 3b)
Is f_directional useful only for streets? Two versions were compared. 3a is Model 2: share-type and sigma unchanged, with the direction score for every candidate. 3b sets the direction score to 0 for buildings and open spaces, so only streets keep it:
\[V_k = w_1 f_\text{distribution} + w_2 f_\text{distance} + w_3 f_\text{dir,street} + w_4\,share_\text{bldg} + w_5\,share_\text{street} + w_6\,share_\text{os} + w_7\,\sigma\]
Both versions have 7 parameters, so they can be compared directly.
| R=10 m | 3a: shared | 3b: street only |
|---|---|---|
| LL(b) | -0.7510 | -0.7520 |
| Accuracy | 0.668 | 0.666 |
| Building / street / open space | 0.872 / 0.362 / 0.398 | 0.872 / 0.335 / 0.414 |
| CV accuracy (grouped) | 0.664 ± 0.046 | 0.658 ± 0.040 |
| Condition number | 117 | 135 |
- Keeping the score only for streets does not help. The fit is a little worse and the model is less stable (the direction score's VIF is 9.26 instead of 7.11).
- It only moves accuracy from street to open space.
- In 3b,
share_bldgbecomes significant (t goes from 0.76 to 2.54). So whethershare_bldgmatters depends on how the direction score is set up.
3a was kept. The same result came back later with the tuned α, in f_directional: shared vs split. Next: Model 4a: alternative-specific parameters. Part of Calibration.