Abstract:
Rooftop farming has become an alternative model to improve urban food security and to make use of urban empty spaces. But the match between surplus supply and consumer demand is inefficient, leading to food waste and lost financial rewards.In this study, we propose a hybrid framework that utilizes both classical graph-theoretic algorithms and GNNs to enhance the surplus-demand matching problem in Dhaka city. A dataset consisting of the geographical, temporal and categorical characteristics of rooftop farming was compiled and studied. Baseline matches were obtained using classical methods such as Maximum Weight Bipartite Matching and link prediction heuristics. Later, GNN models—GCN, GraphSAGE, and GAT—were applied to predict potential links by learning node embeddings that capture spatial, temporal, and relational features. Comparative study based on model predictions, prediction accuracy, precision, recall, F1- score and ROC-AUC suggested that GNN-based models (especially GraphSAGE) achieved better predication performance and flexibility than classical algorithms. The framework also developed recommendations for surplus-demand matching, optimized on the basis of geographical constraints. This study demonstrates how graph-based optimization and machine learning can be combined to enhance food distribution in cities, minimize waste and contribute to sustainability through rooftop food production.
