<![CDATA[Undernourishment constitutes a critical public health challenge in Indonesia with significant impacts on human resource quality and economic productivity. Accurate prediction of undernourishment prevalence is essential for supporting early warning systems and evidence-based food security policy planning. This study conducted comprehensive comparative analysis of three deep learning architectures—Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Transformer—for predicting Prevalence of Undernourishment (PoU) using longitudinal data from 38 Indonesian provinces spanning 2018-2024 with 242 observations. Methodology encompasses systematic preprocessing with minmaxscaler normalization, 70:15:15 dataset split, implementation of three models with hyperparameter tuning via grid search, and evaluation using RMSE, MAE, R², and MAPE. Results demonstrate Transformer achieves superior performance with RMSE 286.02, MAE 217.79, R² 0.8822, and MAPE 48.11%, outperforming GRU (RMSE 315.19, R² 0.8570) and LSTM (RMSE 356.81, R² 0.8167). Learning curves analysis reveals Transformer exhibits faster convergence and smaller training-validation gap (0.075) compared to LSTM (0.10) and GRU (0.105), indicating superior generalization. Although Transformer exhibits higher computational complexity (125,248 parameters), the accuracy-efficiency trade-off remains favorable with inference time of 8.6ms per sample. Transformer superiority stems from its multi-head self-attention mechanism effectively capturing long-term temporal dependencies and complex non-linear patterns. Findings provide evidence-based recommendations for implementing Transformer in food security early warning systems, supporting targeted resource allocation, and contributing to Sustainable Development Goals achievement related to zero hunger.]]>
Copyrights © 2025
