How I actually do the research

The overview is the arc. This is the method. Four habits I can point to in the papers, not a philosophy of science.

1. Start from something a dispatcher or a planner could use

The flood work did not begin as “apply Extra Trees to climate data.” It began as: if water is on the track, what can the railway still run? That produced an operation-based Flood Risk Index — ballast, sleeper, railhead — not a generic hazard score. The buckling work asked which mode of failure you get (snap-through vs progressive), because those are different maintenance problems. The Blue Line GWR asked where land use actually moves ridership, not whether land use “matters” in a citywide regression.

If I cannot say who would change a decision, I do not have a question yet. Consultancy language (“insights,” “frameworks”) is usually a missing user.

2. Geography before the model

I came in through the Geography Olympiad and a four-step model of public transit in Nakhon Ratchasima. That sequence is still the lab order: where, then why people leave, then which mode, then which network. Machine learning sits on top of that, it does not replace it. A transformer that generates trajectories is still a geography problem — agents in space, at hours, with purposes — or it is a sampler with no transport content.

Bangkok is a bad place to skip this. A coefficient estimated on the whole BMA will describe Siam and Lat Krabang with the same sentence. They do not share a sentence.

3. Distrust the global elasticity

The published GWR paper exists because a single land-use coefficient on the MRT Blue Line is a convenient lie. Station catchments are not interchangeable; morning and evening are not interchangeable. I would rather report a map of local relationships that I can argue with than a star on a citywide beta.

That habit is not only for ridership. Flood features that matter in the North may not be the ones that matter on a floodplain in the Central region. An accessibility index that averages Bangkok into one number is a press release. The TRR paper Kongtup led is an index and an application — the application is the geography.

4. The code has to be able to lie to you

I write the pipeline so that a split, a class imbalance, or a leakage shows up as a number I do not like, not as a figure I can still present. SMOTE on the buckling FEM set, ADASYN on the flood years, k-fold instead of a lucky test year: those are not ML fashion. Most railway-flood years are “no flood.” A model that always says zero is accurate and useless.

The next post is that habit in Python and R. This one is the rule: if I cannot break the result from the repo, I do not believe it.

What I do not do

I do not start from a method looking for a case (“we should GWR something”). I do not treat a household survey as ground truth without asking who was home to answer it. I do not put unpublished accuracy on this blog. And I do not pretend a Tokyo foundation model and a Nakhon Ratchasima four-step are the same craft — they share the question, not the estimator.

Papers and orals that this method actually produced are on publications. The thread that connects them is already written. The rest of this series is one paper, one codebase, one in-progress model at a time.