Fold difficulty

Analysis Master 23-07-2026

Cross-validation splits the data into 5 test sets (folds). This table shows the Final model's accuracy in each one:

Fold	Total	Building	Street	Open space
0	0.635	0.896	0.162	0.368
1	0.665	0.906	0.324	0.333
2	0.801	0.897	0.667	0.658
3	0.791	0.877	0.622	0.718
4	0.685	0.906	0.368	0.378

Building accuracy hardly changes (0.877-0.906). Street accuracy changes by more than 4 times (0.162-0.667). Each fold has about the same number of street stays (36-38), so the changes do not come from small numbers. Folds 0 and 1 are the hardest for every type, and also for Random Forest and XGBoost (ML ceiling check, row level). So the difficulty comes from which stays end up in a fold, not from one place type.

Part of Error analysis.

Linked from Confusion matrix · Error analysis · ML ceiling check, row level · RQ2 diagnostics and RQ3 ceiling check