Graph-based multitask learning framework for property price prediction and data imputation in the Sri Lankan real estate market

Property price valuation is important for real estate stakeholders since it provides information about future market movements. However, valuation methodologies in areas such as Sri Lanka confront substantial hurdles due to inadequate datasets and difficulty identifying geographical and structural links. This research presents a novel method for improving property price prediction and data imputation by combining Graph Neural Networks (GNNs) with a multi-task learning (MTL) framework. The methodology uses Graph Attention Networks (GAT) to represent spatial and feature-based relationships inside a property graph, with the "Isolate Training and Validation" strategy used to avoid data leaking and ensure robust generalization. An in-training graph-based imputation module dynamically resolves missing data using neighborhood patterns, whereas the MTL framework optimizes both price prediction and imputation tasks. Evaluated on a real-world Sri Lankan property dataset, the model achieves Mean Squared Error (MSE) values of 0.65 for price prediction and 0.56 for imputation. These findings demonstrate the model's capacity to impute missing data using a local graph structure. However, the price prediction is hampered by significant variability and insufficient capture of broader market trends. To summarize, the graph-based MTL system is a promising AI-driven solution for real estate analytics in emerging economies, with further improvements required to balance job optimization and include global market patterns.