Complexity and transfer
Part of Thesis summary. Research question 3: does a more complex model help, and will the model still work on MILES?
- Overfitting signature: extra complexity improved the fit on training data, but not on unseen data
- ML baseline, stay level (superseded) (replaced later) and ML ceiling check, row level: the model compared with Random Forest and XGBoost
- Sigma in tree importance: why the tree models rank GPS spread (σ) so high
- Raw-density refit (Raw-density alpha profile, Downsampling shrinks the candidate search area): the model refitted without downsampling
Meetings: ML baseline, RQ2 diagnostics and RQ3 ceiling check, Does downsampling matter?.