Scalable Property Valuation Models via Graph-based Deep Learning

TL;DR

This work tackles scalable automated property valuation by modeling geospatial peer dependencies with graph neural networks. It introduces two architectures, PD-GCN and PD-TGCN, that adapt the k-nearest similar house sampling strategy to graph-based learning, with the Transformer Graph Convolutional Network enabling attention-based neighbor weighting. On a large Santiago dataset with final sale prices, both PD-GCN and PD-TGCN outperform several baselines, with PD-TGCN frequently achieving the best overall accuracy (MAPE around 20%). The results highlight the value of attention-driven spatial modeling for real estate appraisal and point to multimodal extensions to further enhance predictive performance in practice.

Abstract

This paper aims to enrich the capabilities of existing deep learning-based automated valuation models through an efficient graph representation of peer dependencies, thus capturing intricate spatial relationships. In particular, we develop two novel graph neural network models that effectively identify sequences of neighboring houses with similar features, employing different message passing algorithms. The first strategy consider standard spatial graph convolutions, while the second one utilizes transformer graph convolutions. This approach confers scalability to the modeling process. The experimental evaluation is conducted using a proprietary dataset comprising approximately 200,000 houses located in Santiago, Chile. We show that employing tailored graph neural networks significantly improves the accuracy of house price prediction, especially when utilizing transformer convolutional message passing layers.

Figures (3)

• Figure 1: PD-GCN diagram
• Figure 2: PD-TGCN diagram
• Figure 3: MAPE by comune for the PD-TGCN method.