Abstract
Geographic information systems’ (GIS) research is widely used within the social and physical sciences and plays a crucial role in the development and implementation by governments of economic, education, environment and transportation policy. While machine learning methods have been applied to GIS datasets, the uptake of powerful deep learning CNN methodologies has been limited in part due to challenges posed by the complex and often poorly structured nature of the data. In this paper, we demonstrate the utility of GCNNs for GIS analysis via a multi-graph hierarchical spatial-filter GCNN network model in the context of GIS systems to predict election outcomes using socio-economic features drawn from the 2016 Australian Census. We report a marked improvement in performance accuracy of Hierarchical GCNNs over benchmark generalised linear models and standard GCNNs, especially in semi-supervised tasks. These results indicate the widespread potential for GIS-GCNN research methods to enrich socio-economic GIS analysis, aiding the social sciences and policy development.
Introduction
Predicting socio-economic outcomes, such as electoral voting patterns, is a fundamental task in social science and policy development. Traditionally, researchers rely on Geographic Information Systems (GIS) data, such as census results. However, applying standard deep learning techniques like Convolutional Neural Networks (CNNs) to GIS data presents significant challenges:
- The Grid Problem: Unlike images with regular pixel grids, GIS data consists of irregular polygons with varying shapes and connectivity.
- The Independence Assumption: Traditional models often treat regions as independent and identically distributed (IID), ignoring the inherent spatial dependencies where neighboring suburbs influence each other.
- The Modifiable Areal Unit Problem (MAUP): Statistical results can vary significantly based on the scale of analysis (e.g., street level vs. suburb level).
This paper introduces a Hierarchical Graph Convolutional Neural Network (GCNN) to address these issues by modeling geography as a graph and capturing multi-scale relationships.
Methodology
The researchers developed a Hierarchical GCNN that mirrors the natural organization of census data.
Graph Construction and Hierarchy
Each geographic region (Statistical Area Level 1, or SA1) is represented as a node in a graph, with edges connecting adjacent regions. To address the MAUP, the model incorporates a multi-level hierarchy, running parallel networks for larger "suburban" areas (SA2).
Information Sharing
The model utilizes "Embedding" and "Projection" operators to facilitate bidirectional information flow between levels. This allows local neighborhoods to learn from broader regional trends. Higher-level information is incorporated via residual connections, ensuring that broader data is only used if it improves local prediction accuracy. Additionally, an adaptive learning metric re-weights neighbor importance based on feature similarity rather than just physical proximity.
Datasets
The study utilized data from the 2016 Australian Federal Election:
- Input Features: Socio-economic data from the 2016 Australian Census, including demographics, income, and education.
- Target Variable: The "Two-Party Preferred" (2PP) vote results.
- Spatial Data: GIS shapefiles defining the adjacency of thousands of regions across Australia.
Results
The Hierarchical GCNN significantly outperformed traditional statistical benchmarks and standard GCNN models across various training data splits.
Key findings include:
- Massive Error Reduction: The model achieved a Root Mean Squared Error (RMSE) of 2.20%, a 70% improvement over the best traditional model (Random Forest, RMSE 7.75%).
- High Explanatory Power: The model reached an $R^2$ of 0.97, explaining nearly all variance in election results using only census and geographic data.
- Robustness in Semi-Supervised Learning: The hierarchical structure acted as a powerful prior, allowing the model to maintain high accuracy even with only 10% training data, outperforming standard GCNNs by up to 14%.
Conclusion
By treating geography as a graph and leveraging hierarchical structures, this research demonstrates a significant advancement in GIS analysis. The Hierarchical GCNN effectively captures spatial dependencies and multi-scale effects that linear models miss. Future work may involve incorporating real-time economic data and applying this framework to urban planning and public health.
Summary updated