| Abstract [eng] |
The aim of this study was to improve the accuracy of real estate price prediction by leveraging both structured data and visual information extracted from property listings. Traditional prediction methods often fall short in accurately estimating real estate prices due to their limited capacity to process diverse data types. To address this issue, a model ensemble was developed, integrating room type classification, property price level estimation, geometric depth information, and a regression model into a unified system. For room type classification, the Vision Transformer (ViT) model with the vit_base_patch16_224 architecture was used, achieving a high accuracy of 97.98%. To predict the price level, the ResNet34 model was implemented, which, despite attaining a lower accuracy of 41.58%, contributed valuable information. Depth metrics were obtained using the Depth-Anything-V2-Metric-Indoor-Large-hf model that aided with depth predictions and the extraction of geometric characteristics from images. The final real estate price prediction was conducted using the XGBoost model. The study's findings indicate that combining structured and visual data enhances prediction accuracy. Using only structured data (H) resulted in RMSE = 74 515.95 and R² = 0.8448. Incorporating price level information (H + L) improved the results to RMSE = 68 109.46 and R² = 0.8703, demonstrating that visual data provides a meaningful refinement to the predictions. These outcomes confirm the significance of visual context, particularly when certain structured data is unavailable or limited. |
