Graph Neural Networks for House Price Prediction: Do or Don't?
_{Margot Geerts, Seppe vanden Broucke, Jochen De Weerdt [2024]}

In this paper, we performed a comprehensive benchmark of graph construction methods and prevalent GNN models for house price prediction. Furthermore, we compare these GNN approaches against an extensive suite of statistical, machine learning, and deep learning models. The results, drawn from six diverse housing datasets, reveal that GNNs are unsuccessful in surpassing machine learning and deep learning baselines. In particular, optimizing the graph structure yields only marginal improvements, with k-nearest neighbor graphs generally exhibiting superior performance. Among the GNN architectures evaluated, GraphSAGE and Transformer-based models demonstrate superior accuracy compared to other GNN variants. Ultimately, the findings suggest a general recommendation against the adoption of GNNs in favor of tree-based models such as LightGBM and CatBoost for house price prediction tasks.

This repository contains all code to reproduce the experiments on the King County dataset.

Repository structure

This repository is organised as follows:

GNNs4HPP
│
├── config/
│    └── data/
│         ├── kc.json: defines the variables for the King County dataset
│         └── network_data_kc.json: defines the graph parameter settings for the different graph construction methods; specific to the KC dataset
│
├── data/
│    ├── raw/
│    │    └── kc_final.csv the King County dataset (from https://www.kaggle.com/datasets/astronautelvis/kc-house-data)
│    └── processed/
│
├── results/
│
├── scripts/
│    ├── main.py: run this to replicate the GNN experiments
│    ├── run_baselines.py: runs the tree-based baseline experiments
│    ├── run_gp.py: runs the GPR baseline experiments
│    ├── run_mlp.py: runs the MLP baseline experiments
│    └── run_pegnn.py: runs the PE-GNN experiments
│
└── src/
     ├── data/
     │    ├── benchmark_dataset.py: defines the MyDataset class
     │    ├── edge_generation.py: defines the functions for graph construction
     │    └── process_kc.py: defines a processing function for the KC dataset
     ├── methods/
     │    ├── model.py: defines the GNN and MLP classes
     │    ├── pegnn.py: defines the classes and functions for PE-GNN
     │    └── utils.py: defines early stopping, training, and evaluation of models
     └── utils/
          ├── dataset_descriptions.py: generates descriptive statistics of all graphs
          ├── feature_attribution.py: calculates feature attribution of XY and hedonic features
          ├── graph_attribution.py: calculates graph attribution with respect to unconnected and random graphs
          └── homophily_measures.py: calculates homophily measures of graphs

Installing

We have provided a requirements.txt file:

pip install -r requirements.txt

The pyg-lib package should be installed additionally, depending on the Torch and CUDA version. This can be done as follows:

pip install pyg-lib==0.3.0 -f https://data.pyg.org/whl/torch-2.1.0+${CUDA}.html

Please use the above in a newly created virtual environment to avoid clashing dependencies.

Usage

After everything is installed properly, we can replicate the experiments in the paper. Firstly, set the project directory in the scripts you want to run. For example, DIR = 'C:\Users\folder\subfolder\GNNs4HPP'

To run the GNN experiments with the King County dataset, kNN graph and GraphSAGE model:

python scripts/main.py --data_name kc --generator knn --gnn_model sage --loader TRUE

To run the tree-based baselines on the King County dataset:

python scripts/run_baselines.py --data_name kc

Citing

Please cite our paper and/or code as follows: Use the BibTeX citation

@article{Geerts2024,
   author = {Margot Geerts and Seppe {vanden Broucke} and Jochen {De Weerdt}},
   doi = {10.1007/s41060-024-00682-y},
   issn = {2364-4168},
   journal = {International Journal of Data Science and Analytics},
   title = {Graph neural networks for house price prediction: do or don’t?},
   url = {https://doi.org/10.1007/s41060-024-00682-y},
   year = {2024},
}