<![CDATA[Poverty measurements based on static indicators often fail to capture the dynamics of economic vulnerability reflected in changes in consumption patterns over time. This study proposes a machine learning framework to identify the consumption vulnerability profiles of rural households in Indonesia using aggregate per capita expenditure data from the Central Statistics Agency (BPS) for the 2013–2024 period at the expenditure stratum level. The main stage of the analysis was conducted using K-Means clustering to form consumption pattern segments, which were then evaluated using internal validation metrics and compared with the Hierarchical Clustering and Gaussian Mixture Model approaches. The K-Means results at K=3 yielded three consumption profiles: STabel, Volatile, and Extreme, with a Silhouette Score of 0.5474, a Davies-Bouldin Index of 0.6471, and a Calinski-Harabasz Score of 291.57. To evaluate the separability of cluster labels, a Random Forest model was used for supervised validation and achieved an accuracy of 96.84% with a macro-F1 of 0.9552 under a stratified cross-validation scheme. SHAP analysis indicated that expenditure structure, particularly the ratio of non-food to food expenditures, was the most contributing feature in distinguishing cluster profiles. These findings suggest that a consumption-pattern-based approach can provide additional insights in economic vulnerability analysis and support the development of proxy simulations for social protection targeting. However, since this study uses aggregate data at the expenditure stratum level, the results are not intended to determine vulnerability or aid recipients at the individual household level without further validation using microdata.]]>
