AI Model Improves Rent Prediction Using Missing Urban Data

This research introduces MARCUS, a novel model designed to improve urban region representation learning, particularly for tasks like real estate appraisal and functional zone identification. A key challenge in this domain is the prevalence of incomplete multimodal urban data. Unlike traditional methods that treat missing data as noise and attempt imputation, MARCUS innovatively views missingness itself as a valuable contextual urban signal. The MARCUS framework operates in three stages. First, "Intra Learning" simultaneously encodes both observed features and the patterns of missing data. Second, "Inter Learning" assesses the reliability of different data modalities to guide how they interact. Finally, a "Fusion" stage employs missing-aware and time-aware gating mechanisms to generate a comprehensive region embedding. Applied to rent prediction, a task influenced by long-term trends and seasonal changes, MARCUS demonstrated state-of-the-art performance on real-world datasets from Sydney and New York. It achieved substantial reductions in Mean Absolute Error (MAE), proving the effectiveness of its missing-aware approach in handling complex urban data.