Re-Defining Vanishing Municipalities in Japan

Kamijima, the island town in Ehime where I grew up and where my college sits, is one of Japan’s “vanishing municipalities”: places whose population is expected to become hard to sustain. The label is decided by population figures alone. Industry and whatever else a town does well play no part in it, even though national statistics on each municipality already exist.

This work, with Keito Inoshita, a doctoral student at Kansai University, asks whether that data can support a broader description of these towns. It belongs to urban computing, the field that uses public data about cities and regions to improve environments and society. The paper, “Re-Defining Vanishing Municipalities in Japan: A Multidimensional Clustering and SLM-Based Policy Insight Framework,” is to appear in the proceedings of the IEEE 15th Global Conference on Consumer Electronics (GCCE 2026), where I will present it in October 2026.

Problem

Japan already holds the data to characterize its municipalities along many axes: SSDSE, a national statistical dataset, and data from the Ministry of Land, Infrastructure, Transport and Tourism (MLIT). Two things shape how it can be used. Existing large language models cannot take a dataset like SSDSE in at once, so the source sits underused. And for a municipality to use such a model at all, it has to be safe enough that leaking information is not a concern.

Method

The pipeline combines existing techniques rather than inventing a new one. Raw tables are cleaned first, removing blanks and unneeded fields. UMAP then reduces the many-dimensional records to two dimensions, and clustering groups the correlated data. That result is used to fine-tune an existing model into a small language model (SLM) that runs on local hardware, so that a municipality could use it without worrying about information leaks.

The aim is a model that knows a specific region well enough to give a grounded answer when asked what that region should do next. Because the data is not being used, there is also no real evaluation index today; a system built on official national data could provide one. The hardest part has been the pace of the field: methods I learned are overwritten almost daily, and the useful skill is seeing how two techniques that look incompatible can be combined.

Status

The paper has not been published yet, so I am not summarizing its results here; I will update this page once it is out. The same idea, a small model fine-tuned on domain data and run entirely on local hardware, is also the basis for the language-model layer I am building into Beacon, my inspection system for dark environments.