A Deep Learning Framework for Personalized Land Parcel Recommendation: Integrating TF-IDF User Profiling with Spatial Analytics

Abstract

Selecting land parcels for purchase or development requires the evaluation of multiple spatial, economic, and user-specific factors. Existing recommender systems in property and spatial decision-support domains commonly rely on collaborative filtering or conventional content-based approaches. However, these methods are often limited in land administration contexts, where transaction records are sparse, parcel characteristics are heterogeneous, and user preferences are context-dependent. To address this limitation, this study proposes an AI-driven land parcel recommendation framework based on a Feature-based Multi-Criteria Recommendation System (FMCRS), integrating user-profile information with spatial and economic land attributes. User identity and professional affiliation were transformed into numerical features using Term Frequency–Inverse Document Frequency (TF-IDF), while land attributes included road width, zoning classification, distance to the central business district (CBD), and predicted land value. The framework was evaluated using data from land valuation professionals in Surabaya City by comparing Ordinary Least Squares (OLS) and Deep Neural Network (DNN) models. The DNN model achieved better predictive performance than OLS, with an R² of 0.92 and a MAPE of 4% on the training dataset, as well as slightly higher testing performance. These results indicate that the proposed framework can better capture non-linear relationships between user profiles and spatial land attributes. Overall, this approach is relevant for supporting personalized, spatially informed, and data-driven land parcel recommendation in complex urban environments with limited historical interaction data.