RQ1 rebuilt with grouped CV
All models were tested again with a fairer method.
- The problem. Three phones recorded each stop at the same time. In the old cross-validation (CV), one phone's recording of a stop could be in the training data and another phone's recording of the same stop in the test data. That made results look better than they really are. All models were fitted again with Grouped cross-validation, which keeps all recordings of one stop together (Model progression).
- What changed. The best model kept the same form and the same accuracy on its training data (0.724). On unseen data it scored 0.709 ± 0.075. The range is wider than before, and more honest.
- What stayed the same. Alpha calibration found α = 2.4297 again. The model with ASC terms was rejected again (Model 5: alternative-specific constants (rejected)), and f_distance² was again shown to be a side effect of the wrong α (The f_distance² artifact). Removing weak terms (Stepwise VIF elimination) gave the 14-parameter model.
- Follow-up.
f_directionaldoes not need a separate weight for each place type (p = 0.067, not significant). The 12-parameter version with one shared weight does better on unseen data (0.716) and is more stable: f_directional: shared vs split → Final model.
The 12-parameter model is the final model of the thesis (Section 5.1.8).
Linked from Assignment framework · Calibration · MNL model development · Thesis summary