SMALL-AREA POPULATION FORECASTING OF SHRINKING CITIES IN SOUTH KOREA - USING SHAP (SHAPLEY ADDITIVE EXPLANATIONS) MACHINE LEARNING

Aug 2, 2026

This study evaluates machine learning for projecting population in small geographic areas of South Korea from 2020 to 2040, comparing results with the cohort-component model. Machine learning shows substantially higher predictive accuracy and reduces forecasting error by incorporating multiple socioeconomic variables beyond birth, death, and migration. The cohort-component approach is expected to generate higher error due to limited explanatory components and difficulty capturing rapid population changes. SHAP interpretation indicates the strongest contributions from pre-population and the variable related to fertile women, with different spatial concentration patterns versus cohort-component outputs.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages22

Reading Time55 min

AccessFree to download

Introduction

Population data and the need for accurate estimation

Population decline and small-area prediction requirements

Purpose and study approach

Theoretical background

Population projection vs forecasting error

SMALL-AREA POPULATION FORECASTING OF SHRINKING CITIES IN SOUTH KOREA: USING SHAP(SHAPLEY ADDITIVE EXPLANATIONS) MACHINE LEARNING (1127) Youhyun Kim1, Donghyun Kim 1* 1 Department of Urban Planning and Engineering, Pusan National University, Busan, Republic of Korea;*[donghyun-kim@pusan.ac.kr](donghyun-kim@pusan.ac.kr) Abstract. The purpose of this study is to identify the utility of machine learning model in projecting the population of small areas. This study was conducted between 2020 and 2040 in the local districts of Korea and compared the research results of cohortcomponent model and machine learning model. As a result of projecting population through the cohort-component method and machine learning, it was identified that the accuracy of the machine learning model was much higher. The cohort-component model is expected to have a high forecasting error because it only explains population change by three component: birth, death, and migration, and it is confirmed that it is almost unpredictable, especially when there are frequent population changes due to new development. On the other hand, the machine learning model reflects various variables such as socioeconomic factors in the population projecting model, which greatly reduces the prediction error. The machine learning model projected that the population would be evenly distributed across the country, especially on the central part of Busan Metropolitan City, while the cohort-component model projected that the population would be concentrated in some areas such as Gijang-gun and Gangseo-gu. The SHAP value interpreted as the machine learning model relying most heavily on the pre-population and fertile women variables to project population. Keywords: Small-area Population Projection, Cohort-Component Method, Machine Learning, SHAP. 1. Introduction Population data is the basis of urban planning and various policy data and is an essential leading indicator. This is because changes in the population structure not only affect all areas such as housing, economy, welfare, and environment, but also determine the size of urban planning facilities and service supply standards based on predicted population data. In other words, population is the most basic data for estimating demand in all areas encompassing cities. Therefore, inaccurate population prediction causes idle capital or congestion costs, causing social inefficiency and degrading the quality of the entire city. This means that a professional understanding of the population structure and size is required for sustainable national land construction, and accurate population estimation should be in the first stage when establishing a basic urban plan. In the era of population decline, questions about the rationality of future population estimates have led to the need for population prediction at a more detailed unit independent of administrative district boundaries. In fact, a lot of research and policies related to population estimation in small regions have recently been discussed domestically and internationally, and research methods that can increase the accuracy of population prediction in small regions are also being devised. (Wilson, 2015; Inoue, 2017) . Statistical techniques used in the existing population estimation process only predict population change trends based on past populations, and there is a limit to setting factors that affect population size and fluctuations. Therefore, the socioeconomic characteristics or regional specificity of the region cannot be considered. Population estimation in small areas should be carried out based on scientific evidence based on big data, away from traditional statistical techniques, as it must ensure the accuracy of high-resolution data. In this context, the grid-based future population prediction model using artificial intelligence techniques can not only diagnose detailed and specific national land phenomena but also reflect nonlinear relationships betwe

FAQs

What is the main goal of this study?

To test whether machine learning can better project small-area population distribution in South Korea and to develop a more suitable forecasting model for small geographic units.

How does the machine learning approach differ from the cohort-component model?

The cohort-component model explains population change using three components (birth, death, and migration), while machine learning incorporates multiple socioeconomic variables, reducing prediction error.

Which factors most influence the model’s population projections according to SHAP?

SHAP interpretation shows the model relies most heavily on pre-population and the variable related to fertile women.

Aug 2, 2026

Upload by Valentina