Predicting Housing Transaction with Common Covariance GNNs

Jinjin Li, Bin Liu, Chengyan Liu, Hongli ZhangIJCAI 2024

Introduction

In the field of real estate data science, research has traditionally focused on price prediction—estimating the market value of properties based on physical attributes and location. However, a significant gap remains in predicting human behavior: specifically, the life trajectory of residents as they transition from renting to first-time homeownership, and eventually to housing upgrades.

Predicting these transitions is inherently difficult due to the high level of "noise" in human decision-making. While some individuals purchase homes for idiosyncratic, unpredictable reasons (specific architectural preferences or personal whims), many others are driven by universal socio-economic factors such as career milestones, marriage, or family expansion. Traditional predictive models often fail because they treat residents as isolated data points and struggle to separate these predictable common drivers from stochastic individual noise.

Methodology: The ccGNN Framework

To address these challenges, this research introduces the Common Covariance Graph Neural Network (ccGNN). The framework is designed to isolate shared economic motivations from individual residuals and leverage the latent relationships between similar residents.

Overview of the ccGNN framework, illustrating (a) factor decomposition into common purchase factors, (b) graph construction using common covariance, and (c) classification of true vs. false positive groups.

The methodology follows a three-step process:

  1. Factor Decomposition: A factor model is employed to decompose the covariance of resident data into two distinct components: common purchase covariance (representing universal drivers) and specific purchase covariance (representing individual noise).
  2. Graph Construction: Rather than relying on explicit social networks, the model constructs a resident-resident correlation graph. In this graph, nodes represent individual residents, and edges represent the similarity of their "common" housing drivers. This captures the latent interconnectedness of people moving along similar life trajectories.
  3. GNN Modeling: The researchers adapted several GNN architectures—including Graph Convolutional Networks (GCN), Graph Attention Networks (GAT), and ChebyNet—by replacing the standard adjacency matrix with the calculated Common Covariance Matrix ($BB^T$). This allows the model to propagate information across "similar" residents to refine classification.

Datasets and Experimental Setup

The study utilizes two distinct micro-datasets from a provincial capital city in China, comprising lease contracts and housing sales records:

  • Dataset 1 (Rent-to-Buy): 25,000 samples tracking the transition from renting to a first-time purchase. Features include age, monthly rent, and rental area.
  • Dataset 2 (Buy-to-Upgrade): 13,000 samples focusing on residents moving from their first home to a second "upgrade" purchase. This dataset includes additional variables such as the price and area of the initial property.

Key Results

The ccGNN approach significantly outperformed traditional machine learning baselines, such as Logistic Regression (LR) and Decision Trees, particularly in terms of sensitivity. In real estate and urban planning, correctly identifying a future buyer (True Positive) is often more critical than overall accuracy.

Performance comparison of various models on the Rent-to-Buy dataset, highlighting the superior sensitivity (TPR) of ccChebyNet.

As shown in the experimental results for the first-time purchase dataset, the ccChebyNet variant achieved a sensitivity (TPR) of 81.2%, outperforming the best traditional baseline (LR) by 14%. For the housing upgrade dataset, the ccTAGCN variant reached a similar sensitivity of 80.9%. Statistical validation using Binomial and Mann-Whitney U tests confirmed that features like age, rental area, and monthly rent were significant drivers for the "True Positive" group.

Discussion and Insights

The core innovation of this work lies in the realization that individual residuals—the idiosyncratic factors unique to one person—actually hinder predictive performance. By mathematically stripping away these personalized factors and focusing exclusively on shared covariance, the model becomes more robust.

Furthermore, the results demonstrate that residents do not need to be socially connected to be "connected" in a predictive sense. If two residents share a similar trajectory in rent inflation and career-age milestones, the GNN can effectively leverage the "neighbor's" data to predict the target's next move.

Conclusion and Future Work

While the ccGNN framework shows high sensitivity in identifying future homebuyers, the authors note that the current data is rooted in specific Chinese demographic trends (the "post-80s and 90s" generations). Future research will explore how well this common covariance structure translates to Western markets with different home-ownership cultures. Additionally, the researchers aim to improve the interpretability of the latent factors and incorporate temporal GNNs to model how these correlations evolve over time.

Summary updated

BibTeX
@inproceedings{li2024predicting,
  author = {Jinjin Li and Bin Liu and Chengyan Liu and Hongli Zhang},
  title = {{Predicting Housing Transaction with Common Covariance GNNs}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2024},
}

Loading related content...

Figure