The fit of military apparel affects soldiers' mobility, protection, and musculoskeletal health, while poor fit inflates logistics costs through excess inventory and exchanges. Conventional sizing relies on a bivariate percentile grid over one or two control dimensions, which cannot reflect multivariate body-shape diversity. This study formulates sizing-system design as a multivariate p-median (facility-location) problem and develops a data-driven, explainable framework using the publicly available 2012 U.S. Army Anthropometric Survey (ANSUR II; 6,068 personnel, 93 measurements). Over a ten-dimension garment-fit space with grading- based tolerances, three size generators—the traditional percentile grid, k-means (centroid prototypes), and p-median/k-medoids (real-body prototypes)—are compared on a size-count versus fit trade-off curve. The data-driven generators dominate the percentile grid: k-means attains with only 15 sizes a higher mean dimensional accommodation than the grid achieves with 30 sizes, and reduces the mean misfit ratio by about 14% at an equal number of sizes, while p-median offers interpretable prototypes that correspond to actual soldiers usable as fit models. A measurement-minimization pipeline then reconstructs the full fit space from a few low-cost measurements using machine learning and builds the sizing system on the reconstructed bodies, evaluating accommodation on the true dimensions to avoid circularity; three measurements (weight, neck circumference, stature) recover 97% of the full-measurement accommodation, and two recover 94.5%. Permutation importance and SHAP (SHapley Additive ex- Planations) identify weight and neck circumference as the most informative measurements. A bootstrap analysis shows that gender-integrated sizing is marginally but significantly better than gender-separated sizing, justifying a unified system for stock-keeping simplicity.

1. 서 론
2. 이론적 배경 및 선행연구
    2.1 사이즈 시스템 설계와 시설입지 최적화
    2.2 인체치수와 사이즈의 관계 및 통제 치수 선정
    2.3 인체측정 데이터 자원
    2.4 머신러닝과 설명가능 인공지능
    2.5 연구의 차별성
3. 연구 방법
    3.1 데이터
    3.2 적합공간과 적합도 지표
    3.3 사이징 생성기
    3.4 측정 최소화 신속 사이징
    3.5 측정 항목 선택(설명가능 측정 설계)
    3.6 성별 통합 분석
4. 분석 결과
    4.1 기술통계
    4.2 RQ1: 사이징 방식 비교 (시설입지 최적화 vs백분위 격자)
    4.3 RQ2-1: 적합공간 복원 성능
    4.4 RQ3: 측정 항목 선택
    4.5 RQ2-2: 측정 최소화 신속 사이징
    4.6 RQ4: 성별 통합 vs 분리 사이징
5. 결론 및 토의
    5.1 결과 요약
    5.2 한국군 적용 함의
    5.3 실무적 함의 및 한국군 적용
    5.4 한계 및 향후 과제
Acknowledgement
References

• Changho Son(Department of Artificial Intelligence and Systems Science, Korea Army Academy at Yeongcheon) | 손창호 (육군3사관학교 인공지능․시스템과학과) Corresponding author