Model 2: share-type and sigma
The second model removes f_temporal. It adds share_type, with a separate weight for each type, and one sigma term:
\[V_k = w_1 f_\text{distribution} + w_2 f_\text{distance} + w_3 f_\text{directional} + w_4\,share_\text{bldg} + w_5\,share_\text{street} + w_6\,share_\text{os} + w_7\,\sigma\]
| R=10 m | R=15 m | |
|---|---|---|
| ρ² vs LL(C) | 0.198 | 0.199 |
| Accuracy | 0.668 | 0.654 |
| Building / street / open space | 0.872 / 0.362 / 0.398 | 0.881 / 0.308 / 0.356 |
| CV accuracy (grouped) | 0.664 ± 0.046 | 0.647 ± 0.042 |
| Condition number | 117 | 98 |
| Term (R=10 m) | w | t |
|---|---|---|
| f_distribution | +3.542 | 8.40 |
| f_distance | -1.209 | -2.60 |
| f_directional | +1.708 | 5.89 |
| share_bldg | +0.219 | 0.76 |
| share_street | -5.552 | -15.68 |
| share_os | -5.887 | -12.79 |
| sigma | +0.475 | 5.87 |
- The model now beats the simple baseline, and accuracy is 15 percentage points higher than Model 1: base.
- Almost all of the gain is for buildings. Building accuracy jumps from about 0.58 to about 0.87. Street stays the same or gets worse, and open space gets worse. The large negative weights
share_streetandshare_ospush the probability toward buildings. This is an early sign of Candidate density bias toward building. - sigma has a clear positive effect. f_directional becomes significant once
f_temporalis removed. - The weight of f_distance now changes sign between 10 m and 15 m, and both values are significant.
Next: Directional scope ablation (3a vs 3b). Part of Calibration.
Linked from Calibration · Directional scope ablation (3a vs 3b) · MNL model development · Model 1: base · Model progression · f_distance · share_type · sigma