Final model
The model the thesis ends with (Section 5.1.8). It has 12 parameters, uses α = 2.4297 and a search distance (R_buffer) of 10 m.
A small \(t\) means the weight is different for each place type (building, street, open space). No \(t\) means one weight shared by all types.
How well it fits. The most important line is CV accuracy: on data it did not learn from, the model picks the correct type for 71.6% of stays.
| LL(0) / LL(C) / LL(b) | -1.0799 / -0.9362 / -0.6557 |
| ρ² vs LL(0) / LL(C) | 0.393 / 0.300 |
| Accuracy (in-sample) | 0.721 |
| CV accuracy (grouped, 5-fold) | 0.716 ± 0.076 |
| CV building / street / open space | 0.896 / 0.429 / 0.491 |
| Condition number | 1,373 |
The weights. w is the learned weight: positive raises the probability, negative lowers it. t shows how clearly the weight differs from zero (about 2 or more is significant). VIF shows overlap with other inputs (above 10 is the usual warning line; see Multicollinearity is structural).
| Term | w | t | VIF |
|---|---|---|---|
| f_distribution, bldg | +1.253 | 2.19 | 5.62 |
| f_distribution, street | +0.913 | 1.57 | 5.37 |
| f_distribution, os | +5.277 | 8.10 | 3.61 |
| f_distance, bldg | -2.102 | -2.04 | 9.37 |
| f_distance, street | -2.293 | -2.51 | 21.42 |
| f_distance, os | -6.554 | -6.14 | 13.05 |
| f_directional | +0.980 | 3.12 | 10.28 |
| share, bldg | -3.152 | -4.64 | 7.90 |
| share, street | -8.088 | -9.57 | 11.95 |
| share, os | -5.873 | -4.17 | 10.13 |
| sigma | +0.778 | 5.89 | 3.14 |
| f_distribution × share_type | +5.779 | 5.66 | 9.81 |
How the model was built: Model progression. Overlap between inputs: Final model correlation matrix. Where it makes mistakes: Confusion matrix. It also holds up without thinning the data: Raw-density refit.
Part of Calibration.
Linked from Calibration · Candidate density bias toward building · Confusion matrix · Contributions · Does downsampling matter? · Final model correlation matrix · Fold difficulty · ML ceiling check, row level · Model progression · Positioning against deterministic assignment · RQ1 rebuilt with grouped CV · RQ2 diagnostics and RQ3 ceiling check · Raw-density refit · Sigma in tree importance · Thesis summary · f_directional: shared vs split · f_distribution · sigma