f_directional: shared vs split
Does f_directional need a separate weight for each place type ("split", the 14-parameter model, \(w_{3,t} f_\text{directional}\))? Or is one weight for all types enough ("shared", 12 parameters, \(w_3 f_\text{directional}\))? Everything else stayed the same.
| Shared (12) | Split (14) | Street only (12) | |
|---|---|---|---|
| Accuracy | 0.721 | 0.724 | 0.723 |
| CV accuracy (grouped) | 0.716 ± 0.076 | 0.709 ± 0.075 | 0.716 ± 0.069 |
| Condition number | 1,373 | 2,430 | 1,431 |
Split vs shared: LR = 5.41, df = 2, p = 0.067. This is not significant: the two extra weights do not improve the fit enough. They raise training accuracy by 0.3 percentage points, but lower accuracy on unseen data and double the condition number. That is overfitting (Overfitting signature).
The shared weight is clearly significant (t = +3.12). In the split version, only the building weight passes 1.96. "Street only" performs about the same as shared at this size, and never better. In the 7-parameter models, shared wins on every measure (CV 0.666 against 0.656), the same result as in Directional scope ablation (3a vs 3b).
Decision: shared, 12 parameters → Final model, the final model of the thesis (Section 5.1.8). Part of Calibration.