Earticle

초록

영어

This study proposes a machine learning model to predict the priority of vacant house occurrence at the administrative-dong level, using Yeosu City, Jeollanam-do, as a case study, against the backdrop of growing social costs associated with vacant houses inrural areas and small-to-medium regional cities experiencing population decline and aging. A dataset covering 26 administrative-dong units was constructed by integrating building registry records (48,529 entries), administrative-dong-level vacant house statistics, population and household data at the administrative-dong and administrative-ri levels, elderly population statistics (aged 65 and over), and single-person household data. A three-tier priority label (low, medium, high) was generated based on the distribution of vacancy ratios. The prediction model was designed as a LightGBM-based two-step classification framework: Step 1 distinguishes areas with a vacancy tendency from those without, while Step 2 differentiates medium- from high-priority areas. Single-split cross-validation yielded F1 scores of 0.765 (Step 1) and 0.900 (Step 2), and repeated stratified cross-validation (20 repeats × 5-fold) confirmed stable performance, with mean accuracies of 0.694 (SD = 0.054) for Step 1 and 0.791 (SD = 0.090) for Step 2. SHAP analysis identified the proportion of structurally vulnerable buildings (block and wood construction), average lot size, and building age-related variables as the most influential predictors of vacant house priority. The elderly population ratio, computed directly from raw administrative records on the population aged 65 and over and single-person households, demonstrated meaningful explanatory power, ranking seventh with a SHAP contribution of 0.501. The prediction results showed that older urban-core administrative-dong units exhibit higher vacant house risk than rural eup and myeon areas. Notably, Hwajeong-myeon and Nam-myeon - which recorded the highest elderly population ratios in Yeosu City - were classified as medium or low priority, suggesting that population aging alone is not a sufficient condition for vacant house occurrence and becomes a risk factor only when combined with physical building deterioration.

목차

Abstract
1. 서론
1.1 연구의 배경 및 필요성
1.2 연구의 목적 및 범위
2. 선행 연구 및 대상지 특성
2.1 선행연구
2.2 연구 대상지 특성
2.3 자료 구성
3. 연구 방법
3.1 특성공학(Feature Engineering)
3.2 우선순위 라벨 생성
3.3 2단계 LightGBM 분류 모델
3.4 하이퍼파라미터 최적화
3.5 모델 평가 및 반복 교차검증
4. 분석 결과
4.1 행정구역별 빈집 우선순위 분류 결과
4.2 모델 성능 평가
4.3 반복 교차검증을 통한 안정성 검증
4.4 SHAP을 통한 변수 중요도 해석
5. 결론
참고문헌

키워드

저자

• 정금호 [ Chung, Kumho | 전남대학교 건축디자인학과 교수, 공학박사 ]

참고문헌

발행기관

간행물

표지보기 관심저널 등록

• 간행물명

  한국농촌건축학회논문집 [Journal of the Korean institute of rural architecture]

• 간기

  계간

• pISSN

  1229-2532

• 수록기간

  1999~2026

• 등재여부

  KCI 등재

• 십진분류

  KDC 522 DDC 630