Paper Type
Complete
Paper Number
PACIS2026-1848
Description
Smart city indices embedded in digital benchmarking platforms increasingly shape urban governance decisions, yet criteria weighting, a critical algorithmic design choice, remains underexamined. This study compares two paradigms for deriving criteria weights: objective multi-criteria decision-making (MCDM) methods and machine learning (ML) feature-importance approaches, applied to the IMD Smart City Index 2024 dataset covering 45 Asia-Pacific cities and 39 indicators. TOPSIS rankings are generated under weight configurations from four MCDM and four ML methods. Results reveal systematically different weight structures: Gini coefficients of 0.10–0.37 for MCDM versus 0.42–0.50 for tree-based ML methods, with ML approaches concentrating importance on digital service indicators and allocating 67–75% of weight to the Technology pillar. Despite these divergences, city rankings remain highly correlated, suggesting the disagreement is one of interpretability rather than decision outcomes. The study contributes a comparison framework for designing transparent, configurable IS artifacts for urban governance.
Recommended Citation
Kizielewicz, Bartłomiej; Bączkiewicz, Aleksandra; Wątróbski, Jarosław; and Sałabun, Wojciech, "Comparing Machine Learning and MCDM Weighting Paradigms in Smart City Benchmarking" (2026). PACIS 2026 Proceedings. 13.
https://aisel.aisnet.org/pacis2026/iot_smartcity/iot_smartcity/13
Comparing Machine Learning and MCDM Weighting Paradigms in Smart City Benchmarking
Smart city indices embedded in digital benchmarking platforms increasingly shape urban governance decisions, yet criteria weighting, a critical algorithmic design choice, remains underexamined. This study compares two paradigms for deriving criteria weights: objective multi-criteria decision-making (MCDM) methods and machine learning (ML) feature-importance approaches, applied to the IMD Smart City Index 2024 dataset covering 45 Asia-Pacific cities and 39 indicators. TOPSIS rankings are generated under weight configurations from four MCDM and four ML methods. Results reveal systematically different weight structures: Gini coefficients of 0.10–0.37 for MCDM versus 0.42–0.50 for tree-based ML methods, with ML approaches concentrating importance on digital service indicators and allocating 67–75% of weight to the Technology pillar. Despite these divergences, city rankings remain highly correlated, suggesting the disagreement is one of interpretability rather than decision outcomes. The study contributes a comparison framework for designing transparent, configurable IS artifacts for urban governance.
Comments
10-IoT