Thesis summary

Thread Master 05-08-2026

Beyond the Building: A Probabilistic Framework for GPS Assignment Under Positional Uncertainty

When a phone records GPS points while someone stays in one place, the points do not tell you exactly where the person was. In a city, the GPS error is often larger than the distance between a building, the street in front of it, and the open space next to it.

Most studies solve this by picking the nearest building. This thesis does something different. For each stay, it gives a probability for building, street and open space, so the uncertainty stays visible. The model learns how to do this from a field survey in Hiroshima, where participants recorded where they really were.

New to the terms? Start with the Glossary.

Why and what

• Why "nearest building" is not good enough: Motivation
• The three research questions: Research questions

Threads

The notes are grouped by theme, not by thesis chapter:

1. Problem and positioning: the problem, the research questions, and how this work compares with earlier methods
2. Ground truth data: the Hiroshima map layers and the 907 stays with known answers
3. Assignment framework: how the method works: candidates, four scores, and a model that turns scores into probabilities
4. Calibration: research question 1, how the final 12-parameter model was built step by step
5. Error analysis: research question 2, where the model makes mistakes and why streets are often mistaken for buildings
6. Complexity and transfer: research question 3, whether more complex models help, and whether the model will work on new data
7. Outlook: what the thesis contributes, its limits, and next steps

Where it stands

• The final model has 12 parameters. On data it had not seen, it picks the correct type for 71.6% of stays (± 7.6%), written as grouped CV accuracy 0.716 ± 0.076 (Final model).
• It finds 89.6% of building stays, but only 42.7% of street stays and 49.2% of open-space stays. When it gets streets or open spaces wrong, it usually says "building" (Confusion matrix, Candidate density bias toward building).
• Random Forest and XGBoost fit the training data better, but on unseen data this simpler model does better (ML ceiling check, row level).
• The thesis is complete. The questions that were still open in these notes are answered in the final text (Final model, Ground truth cleaning, Tension with the street-space argument).

Dated updates

MNL model development · ML baseline · RQ1 rebuilt with grouped CV · RQ2 diagnostics and RQ3 ceiling check · Does downsampling matter?

Linked from Assignment framework · Calibration · Complexity and transfer · Error analysis · Glossary · Ground truth data · Outlook · Problem and positioning