Deep Embedded Clustering for Indonesian Protein, Fat, and Energy Availability Data

Abstrak

Understanding disparities in regional food availability is crucial for food security policies. Most previous studies on Indonesian food availability use conventional clustering methods. These methods operate directly on the feature space and may miss complex, non-linear relationships in nutritional data. This limitation highlights the need for advanced analytical approaches to uncover deeper patterns. This study analyzes patterns of provincial food availability in Indonesia using Deep Embedded Clustering (DEC). It uses per capita indicators of energy, fat, and protein from both plant and animal sources, as well as the 2023 Food Consumption Pattern (FCP) score. DEC integrates representation learning with clustering. This allows the model to capture latent structures and nonlinear relationships that traditional clustering cannot identify. The analysis began by comparing K-Means and Hierarchical Clustering using the silhouette score to generate pseudo-labels for the DEC model. Hierarchical Clustering with Ward linkage and Euclidean distance achieved the highest silhouette score (0.3958) and was used for pseudo-label generation. Two DEC configurations were implemented, showing improved clustering performance. These achieved silhouette scores of 0.7829 (DEC-1) and 0.6385 (DEC-2). The results reveal four distinct clusters of Indonesian provinces, each with different food availability characteristics. These range from balanced, nutrient-rich regions to provinces with more limited or specific nutritional patterns. The findings show that DEC can capture complex structures in nutritional data. It produces more meaningful clusters than conventional approaches. In practice, the identified clusters provide policymakers, nutrition experts, and the food industry with useful insights for region-specific strategies. These strategies can improve food security and nutritional balance. Theoretically, this study contributes to the use of deep learning-based clustering in food availability analysis. It is especially relevant in national food security research. Future research may extend this approach by integrating time-series data and spatial analysis. This will help understand the temporal and regional dynamics of food availability in Indonesia.

Kata kunci

• Deep Learning
• Auto Encoder
• Hierarchical Clustering
• Data Analysis
• Deep Embedding Clustering