Does downsampling matter?
Sensei's question. Before fitting, the survey data is thinned out to look like MILES (downsampling). Does this step change the result? To check, the Final model was fitted again on the full survey data, without thinning.
- Downsampling does no harm. On unseen data, the thinned data gives 0.716 ± 0.076 and the full data gives 0.697 ± 0.062. The model is also more stable with thinned data (condition number 1,373 vs 2,865) (Raw-density refit).
- α changes. With the full data, the best α is 4.6159 instead of 2.4297. But the fit changes very little over a wide range of α, so this value is less certain (Raw-density alpha profile).
- Side finding. Downsampling also makes the search area around each stay smaller, so each stay gets fewer candidates (Downsampling shrinks the candidate search area).
Linked from Complexity and transfer · Ground truth data · Thesis summary