RQ2 diagnostics and RQ3 ceiling check
Checks on the Final model. The model was not fitted again; these checks only look at its results.
- RQ2, where the model fails. The model finds 89.6% of building stays, but only 42.7% of street stays and 49.2% of open-space stays. When it gets a street or open space wrong, it usually says "building" (Confusion matrix). The reason: street stays have the most street candidates around them, and the
share_typeterm punishes that (Candidate density bias toward building). The difficult test sets are difficult for all three types (Fold difficulty). - RQ3, could another model do better? On unseen data, the simple model beats Random Forest and XGBoost (0.716 vs 0.699 and 0.695), even though both fit the training data better (ML ceiling check, row level). This reverses the result from ML baseline.
- Appendices. The thesis appendices were put together from existing data: Final model correlation matrix, Self-report form, the five experiment groups, and Ground truth drop list. One count did not match, and the thesis has since been corrected: Ground truth cleaning.
- Discussion draft. A four-part draft of the discussion, which became the Discussion section of the thesis: Problem and positioning and Error analysis.
Linked from Complexity and transfer · Error analysis · Ground truth data · ML baseline · Problem and positioning · Thesis summary