A multi‑view contrastive learning framework for spatial embeddings in risk modelling
This paper addresses the difficulty of 𝗶𝗻𝘁𝗲𝗴𝗿𝗮𝘁𝗶𝗻𝗴 𝗰𝗼𝗺𝗽𝗹𝗲𝘅, 𝗵𝗶𝗴𝗵-𝗱𝗶𝗺𝗲𝗻𝘀𝗶𝗼𝗻𝗮𝗹 𝘀𝗽𝗮𝘁𝗶𝗮𝗹 𝗱𝗮𝘁𝗮, 𝘀𝘂𝗰𝗵 𝗮𝘀 𝗰𝗹𝗶𝗺𝗮𝘁𝗲 𝗮𝗻𝗱 𝘀𝗮𝘁𝗲𝗹𝗹𝗶𝘁𝗲 𝗶𝗺𝗮𝗴𝗲𝗿𝘆, 𝗶𝗻𝘁𝗼 𝗽𝗿𝗲𝗱𝗶𝗰𝘁𝗶𝘃𝗲 𝗺𝗼𝗱𝗲𝗹𝘀 𝗳𝗼𝗿 𝗶𝗻𝘀𝘂𝗿𝗮𝗻𝗰𝗲.
The study proposes a novel multi-view contrastive learning framework designed to generate low-dimensional spatial embeddings. This method aligns data from multiple sources (e.g., satellite imagery and OpenStreetMap features) with coordinate-based encodings.
The resulting embeddings are shown to consistently improve predictive accuracy in risk models, demonstrated through a case study on French real estate prices. The paper highlights that the embeddings capture spatial structure, enhance model interpretability, and exhibit transferability to unobserved regions.
The study proposes a novel multi-view contrastive learning framework designed to generate low-dimensional spatial embeddings. This method aligns data from multiple sources (e.g., satellite imagery and OpenStreetMap features) with coordinate-based encodings.
The resulting embeddings are shown to consistently improve predictive accuracy in risk models, demonstrated through a case study on French real estate prices. The paper highlights that the embeddings capture spatial structure, enhance model interpretability, and exhibit transferability to unobserved regions.