Candidate density bias toward building
Why the model often says "building" when the correct answer is street or open space.
A first guess would be that street stays have too few street candidates. The data shows the opposite. Stays that really were on a street have the most candidates of their own type:
| True type | Same-type candidates, mean | median | Total candidates, mean |
|---|---|---|---|
| Building | 7.09 | 5 | 22.99 |
| Street | 13.17 | 11 | 36.61 |
| Open space | 5.42 | 3 | 18.34 |
The share_type term gives a lower value to a candidate with many neighbours of the same type (the weight share_street is -8.09 in the Final model). The map stores streets as many short pieces next to each other (Spatial entity layers). So a stay on a street is punished for a crowd of street pieces that the map itself created. The probability then moves to the nearest building (Confusion matrix).
This comes from the street map, so it can be fixed. The GPS data itself does not set the limit (Contributions). But it weakens the argument that streets are real places of activity: Tension with the street-space argument, Candidate density as boundary-matching difficulty.
Part of Error analysis.