Giving Eyes to Automated Valuation Models: Assessing the Explanatory Power of Image-Based Condition Variables Using Machine Learning
Abstract
Current Automated Valuation Models (AVMs) for real estate price appraisals often overlook qualitative factors, such as property condition, which can significantly impact valuation performance. This study utilizes two datasets, comprising 8,865 and 10,486 apartments respectively, to explore how incorporating image-based condition variables can enhance AVMs for apartment valuation in Oslo, Norway. By integrating images, the AVMs are effectively given "eyes," allowing them to account for the visual condition of apartments. To achieve this, a convolutional neural network (CNN) was trained on 250,000 images to classify 428,000 images into four room types: bathroom, bedroom, kitchen, and living room. Additionally, four separate CNNs were developed to evaluate the condition levels of these room types on a scale from one to five. Condition variables, including the average condition of the four rooms, were incorporated into hedonic linear regression, XGBoost, and multi-layer perceptron AVMs, which were trained both with and without these variables for comparison. Their impacts were analyzed using explainable artificial intelligence tools (XAI). The results demonstrated that including condition variables improved predictive performance across all AVMs. XGBoost outperformed the others, achieving relative improvements in MSE and MAPE of 11.5% and 4.34%, respectively, with corresponding reductions in error metrics of 0.13% and 0.45%. A near-linear relationship was observed between price and the average condition level, which had the greatest overall impact on valuation for the condition variables and was among the most important features. Among room types, bathrooms and kitchens had the strongest influence on valuation, followed by bedrooms and living rooms. These findings underscore the importance of image-based condition variables in enhancing AVM performance and offer valuable insights for researchers, appraisers, homeowners, and investors alike.