MNL model development
The first time the model was fitted to the cleaned ground truth (907 stays with known answers). MNL is the type of model used here: it turns scores into probabilities (see Glossary).
- Search distance. A search distance (
R_buffer) of 10 m works better than 15 m for every model tried: R_buffer: 10 m vs 15 m. - Six models compared, each adding something to the one before: Model 1: base → Model 2: share-type and sigma → Directional scope ablation (3a vs 3b) → Model 4a: alternative-specific parameters → Model 5: alternative-specific constants (rejected) → Interaction term screen (α = 1). Accuracy (the share of stays with the correct type) rose from 0.518 to about 0.724.
- α. The setting α controls how fast the distance score drops with distance. It was tuned to about 2.4 on Model 4a (Alpha calibration). With α at 2.4, the extra term f_distance² was no longer useful and was removed (The f_distance² artifact).
- Best model at that time. Separate weights for each place type, plus one combined term (f_distribution × share_type). 14 parameters, accuracy 0.724.
- High VIF values. Some important terms have high VIF values (a warning sign that terms overlap). Here that comes from how the model is built, and it is not a mistake: Multicollinearity is structural.
Later corrected by RQ1 rebuilt with grouped CV.
Linked from Assignment framework · Calibration · Thesis summary