Research

I extend this course project with an independent study. First in my line of work on agents as models of how people decide and act.

The question

The course tabs build the standard ML models on this data. The research asks a additional question : when does an agent built from a person’s ordinary days stop being a reliable description of that person, and how much fresh evidence brings it back? YJMob100K makes this testable, because it follows the same 25,000 people through 60 ordinary days and 15 days of an emergency.

Four studies

Study 1, place semantics (supporting). Foursquare check-ins carry real place types, so I can test how well the kind of place is predicted from context alone: the cell, the hour, the weekday and the person’s history. The classifier tabs of the course do this work. A second, exploratory step asks whether the 85 anonymous place categories of YJMob100K can be recovered from the daily rhythm of each cell, validated first on Foursquare with the names hidden. Any recovered label will be reported as inferred, never as a feature of the dataset.

Study 2, behaviour change (supporting). Person profiles from the ordinary days, clustering and PCA on the course tabs, and a plain description of what changed between days 45 to 59 and 60 to 74. The observation rate is reported as an outcome in its own right, since being recorded at all changes with the emergency.

Study 3, reliability (core). Simple agents of each person’s movement, from a last-cell rule to a learned mix of long-run and recent habits, trained on days 0 to 44, validated on 45 to 59 and frozen. Each agent is tested on days 67 to 74 with 0, 1, 3 or 7 days of post-change data, under equal information, so the update mechanism is a thing that differs. Context contains: movement history alone, history plus the anonymous place counts, and history plus the Study 1 semantics if its validation passes. Outcomes are next-position error, short rollouts of 3 and 7 steps, and improvement over the frozen agent, all by pre-change behaviour group. Two hypotheses: ordinary-period validation does not fully predict post-change reliability, and fresh evidence helps with diminishing returns. Both can fail.

Study 4, a learned update schedule (RL extension). Study 3 fixes when an agent is updated. Study 4 lets an agent learn it. Under a budget of k updates, an agent sees a small state each day, yesterday’s error, its trend, updates left and the weekly phase, and chooses to keep or to update. The reward is the negative prediction error, scored by replaying the recorded observations. Tabular Q-learning written from scratch in NumPy, trained on pre-emergency days only, frozen before day 60, and compared under the same budget with never updating, updating on the first k days, the most recent k days, and a threshold rule. If the threshold rule is enough, that is the finding.

Progress

Updated October 2026, after Part 2 of the course project.

Reading anonymous places through named ones

YJMob100K counts 85 categories of places in every cell but does not say what the categories are, so a cell full of category 59 could be offices, homes or bars. Study 1 asks whether the kind of place can be read from when people are there. The first step was done this month.

• The rhythm of a cell. For every cell with at least 500 records on days 0-59, the share of its records in each half hour, on workdays and on rest days separately. The size of the crowd is divided out, so only the timing remains.
• Five rhythm types. K-means sorted the 11,858 cells into five types, named by the hour at which they are busiest on a workday: 8 am, noon, 5 pm, 6 pm and 8 pm (the Clustering tab shows them). The cells form a gradient rather than separate groups, so the types describe the range of rhythms more than they find natural kinds.
• A named mirror. Foursquare check-ins from Tokyo carry the name of the place, folded into eight groups such as Transit, Food or Work and Institutions. The same rhythm was built for each group.
• The comparison. Each side minus its own average rhythm, then the cosine similarity between every YJMob type and every Foursquare group. Without subtracting the average every pair looks alike (0.42 to 0.96), because all places share the basic shape of a day.

Figure 1. YJMob place rhythms against Foursquare place groupsClose

What came out is a consistent picture. The noon cells and the 8 am cells look like Work and Institutions (0.83 and 0.71), the 5 pm cells like Shopping and Services (0.54), the 6 pm cells like Nightlife (0.88) and the 8 pm cells like Home (0.76). The busiest-at-noon type is also the one that lost the most records in the emergency, 15% on workdays, which fits a workplace that fewer people came to.

What it does not show is just as important. A rhythm says when people are in a cell, not what the cell contains, and the two sources record people differently: YJMob records a phone, Foursquare a person who chose to check in, which is why evening leisure and home look so alike here. The check-ins come from another provider and from 2012-13, and the two grids are never aligned on a map. The labels above are inferred and reported as such; they are not names the dataset contains, and no attempt is made to identify the city.

Next comes the harder step of the study: estimating a rhythm for each of the 85 categories separately, by treating a cell’s rhythm as a mix of the rhythms of the places it holds, and testing the method first on Foursquare with the names hidden, where the right answer is known.

Signals for an agent that updates itself

Study 4 asks whether an agent can learn when to update its picture of a person. That needs a signal it could watch. The PCA tab measured two on real data: how far a person’s fortnight sits from the centre of their own kind of mover, and how much of it the five frozen components cannot reconstruct. Both grow in the emergency (median 2.81 to 3.15 and 1.65 to 1.80) and grow most for regular commuters. They will be among the candidate inputs when the update agent is built in November.

Limits

One emergency in one undisclosed city, so nothing transfers to emergencies in general. The study does not try to identify the city or any person, and it does not align the grid with a map. Results that go against the hypotheses will be reported like any others.