Code for the Rethinking Real Estate Pricing with Transformer Graph Neural Networks (T-GNN)
Abstract—Real Estate Pricing (Appraisal) seeks to estimate the “true” value of a house in the current marketplace. This task presents several distinct challenges that makes the development of machine learning solutions a non-trivial task. First, unlike for several other asset classes, the notion of an intrinsic value for a house is vague, producing unique challenges for interpretability. Second, the real estate market is not as efficient or as liquid as other asset classes (stocks, bonds, etc.). So, the available data is often sparse, with highly irregular temporal and spatial latency. Finally, real estate prices are highly dependent on both temporal and geographic changes (when and where). In this paper, to address these challenges, we propose a novel Transformers Graph Neural Net (T-GNN) framework. Specifically, we provide an end-to-end, data/preprocessing agnostic solution with T-GNN that leverages an intuitive representation of real estate transactions as a graph. (1) We find that the T-GNN significantly boosts the performance (40%-50%) across all metrics compared to tabular-based, non-GNN solutions (e.g., XGBoost). (2) We propose and evaluate a novel interpretability scheme to generate a list of comparable homes for each house valuation, improving on the baseline method by 50%. (3) We investigate why T-GNN significantly outperforms the respective benchmarks, attributing the superior performance to a better representational understanding of time and space.