ML baseline
Sensei's question: could machine-learning models such as Random Forest or XGBoost do much better? If yes, that would show how far the simple model is from the best possible result.
- Result at the time. Random Forest did a little better (+3.8 percentage points in cross-validation, CV: testing on data the model did not learn from). XGBoost was about the same as the simple model. All tree models scored 100% on their own training data, which means they memorized it: ML baseline, stay level (superseded).
- GPS spread. σ (how spread out the GPS points are) was one of the most important inputs for both tree models, even though it has only a small weight in the simple model: Sigma in tree importance.
- Decision. The tree models were not adopted, because they cannot give a clear probability for each place.
This comparison was later replaced by a fairer one in RQ2 diagnostics and RQ3 ceiling check.