Catch-up summary

The master's thesis is complete. Two things are in progress now:

• A paper for the journal Computers, Environment and Urban Systems (CEUS), in drafting. It presents the method from the thesis: for each GPS stay, it gives a probability that the person was in a building, on a street, or in an open space, instead of picking one place. The model learns from 907 stays with known answers from a field survey in Hiroshima. On data it has not seen, it picks the correct type for 71.6% of stays (grouped CV accuracy 0.716 ± 0.076), and it does better there than Random Forest and XGBoost. Its main weakness, mistaking streets and open spaces for buildings, comes from how the street map is cut into many short pieces, not from the GPS data itself. Start from Thesis summary.
• A poster and a presentation for CSIS DAYS 2026, in drafting.

Threads

Hubs by theme: the thesis summary, then one list of notes per theme.

• Thesis summary
• Problem and positioning
• Outlook
• Ground truth data
• Error analysis

All 8

Progress

Dated updates for sensei, and slide exports.

• PhD interview slides PDF
• Does downsampling matter?
• RQ2 diagnostics and RQ3 ceiling check
• RQ1 rebuilt with grouped CV
• ML baseline

All 6

Notes

Small pieces of progress: method, data, interpretation, discussion.

• Glossary
• Study area
• Spatial entity layers
• Research questions
• Overfitting signature

All 34

Analysis

Raw analysis output: tables, numbers and diagnostics.

• Raw-density refit
• Raw-density alpha profile
• Downsampling shrinks the candidate search area
• Sigma in tree importance
• ML ceiling check, row level

All 25

Inbox

Open items not yet settled with sensei.

Nothing yet.

About this research

Hafiyyan Hilmy Fawwaz · Urban Transportation and Planning Lab, Hiroshima University

A GPS staypoint rarely tells you which place a person actually visited: positional error is often larger than the gap between a building, the street in front of it and the open space beside it. This research builds a probabilistic framework that assigns each staypoint to an urban spatial entity (building, street or open space) under explicit positional uncertainty, instead of snapping it to the nearest one.

Dates are written DD-MM-YYYY.