🤖 AI

Heterogeneous Graph Neural Networks with Post-hoc Explanations for Multi-modal and Explainable Land Use Inference

This paper proposes an explainable framework that integrates Heterogeneous Graph Neural Networks with post-hoc Explainable AI techniques to improve the accuracy and interpretability of multi-modal urban land use inference, thereby supporting data-driven city planning and policy-making.

Original authors: Xuehao Zhai, Junqi Jiang, Adam Dejl, Antonio Rago, Fangce Guo, Francesca Toni, Aruna Sivakumar

Published 2026-05-11

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Xuehao Zhai, Junqi Jiang, Adam Dejl, Antonio Rago, Fangce Guo, Francesca Toni, Aruna Sivakumar

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a city as a giant, living puzzle. For decades, city planners have tried to figure out what each piece of the puzzle is used for (is it a home? a shop? a park?) by looking at static maps or asking people surveys. But cities are messy, dynamic, and often don't match the official plans.

This paper introduces a new, smarter way to solve that puzzle using data from how people move around (like bus, subway, and bike usage) combined with AI that understands connections.

Here is a breakdown of their approach, using simple analogies:

1. The Problem: The "Isolated Room" Mistake

Previous attempts to guess land use were like trying to understand a person's job by only looking at them in a single room, ignoring who they talk to or where they go next.

The Issue: Old methods treated every location as if it existed in a vacuum. They didn't realize that a bus stop is connected to a subway station, or that a bike share spot is linked to a nearby park. They also ignored that different types of transport (buses vs. bikes) tell different stories.
The Black Box: Even when AI worked well, it was a "black box." It gave an answer but couldn't explain why. City planners need to know the "why" to trust the AI with long-term policy decisions.

2. The Solution: The "Heterogeneous Graph" (The City Web)

The authors built a digital model called a Heterogeneous Graph Neural Network (HGN).

The Analogy: Think of the city not as a list of addresses, but as a giant, multi-layered spiderweb.
- The Nodes (Spiders): These are the stops (bus stops, subway stations, bike docks).
- The Edges (Strings): These are the connections. Some strings are roads, some are train tracks, and some are walking paths between different types of stops.
- The "Heterogeneous" Part: This is the key. In a normal web, all spiders look the same. In this web, some spiders are blue (subway), some are green (bus), and some are orange (bikes). The AI is smart enough to know that a blue spider behaves differently than a green one, but they are still part of the same web.
The Input: They fed the AI data on how many people got on and off these vehicles every 15 minutes throughout a day.
The Output: The AI predicts the "intensity" of different land uses (e.g., how much "office" activity vs. "residential" activity is happening there).

3. The Results: A Better Map

When they tested this "City Web" AI against older, simpler models:

It won every time. It was significantly more accurate at guessing land use, especially for tricky categories like "sustenance" (restaurants/cafes) and "office."
It fixed the "blind spots." Older models tended to guess the same thing for whole neighborhoods (like saying an entire area is just "residential"). The new model saw the nuances, realizing that one street might be busy with workers while the next block is full of shoppers.

4. The "Why" (Explainability)

The most important part of this paper is that they didn't just build a black box; they built a transparent one. They used two "flashlights" to show how the AI thinks:

A. The "Feature Attribution" Flashlight

This asks: "Which part of the daily schedule mattered most for this prediction?"

The Finding: The AI's logic matched human intuition perfectly.
- For Offices, the AI saw that the morning rush (people arriving) and the evening rush (people leaving) were the biggest clues.
- For Homes, it saw the opposite: people leaving in the morning and returning in the evening.
- For Leisure/Shopping, the AI realized these places weren't about the morning rush but were busy in the evenings.
Why it matters: Because the AI's "reasoning" matched what human experts already knew, planners can trust its predictions.

B. The "Counterfactual" Flashlight

This asks: "What is the smallest change needed to turn this area into a 'perfectly mixed' community?"

The Analogy: Imagine a neighborhood that is currently 90% offices. The AI acts like a "What-If" simulator. It says, "If we just changed the type of road nearby, or if the bike usage patterns shifted slightly, this area could become a balanced mix of homes and shops."
The Finding: They found that for some areas, changing the traffic patterns (node features) was the key. For others, changing the type of road (edge types) was the most important factor. This helps planners understand exactly what levers to pull to fix a neighborhood.

Summary

The authors created a system that treats the city like a complex, multi-colored web of connections rather than a list of isolated spots. By using this "City Web" AI, they can predict land use more accurately than ever before. Crucially, they added a "translator" that explains the AI's logic in plain English, proving that the AI isn't just guessing—it's actually understanding the rhythm of city life.

The Bottom Line: This tool helps city planners see the city as it really is (dynamic and connected) rather than how it was planned on paper, and it gives them the evidence they need to make better, more trusted decisions.

Technical Summary: Heterogeneous Graph Neural Networks with Post-hoc Explanations for Multi-modal and Explainable Land Use Inference

Problem Statement
Urban land use inference is critical for city planning and policy-making but faces significant challenges in accuracy, data integration, and interpretability. Traditional methods often rely on simplified assumptions or survey data, failing to capture the complex, dynamic nature of urban environments. While recent data-driven approaches utilize multi-modal mobility data (e.g., mobile phone records, GPS trajectories), existing studies frequently process samples independently, ignoring spatial correlations between neighboring objects and the heterogeneity among different transport services. Furthermore, the "black-box" nature of deep learning models hinders their adoption in policy contexts where transparency and the ability to extrapolate for long-term decisions are essential.

Methodology
The authors propose an explainable framework that synergizes Heterogeneous Graph Neural Networks (HGNs) with post-hoc Explainable AI (XAI) techniques. The methodology is structured into three main components:

Heterogeneous Graph Construction:
The urban mobility system is modeled as a heterogeneous graph where nodes represent different types of mobility facilities (tube stations, bus stops, bike-sharing stations) and edges represent spatial connections (road networks, transit routes).
- Nodes: Features are derived from daily passenger inbound data (15-minute intervals from 06:00 to 22:00) for three modes: Tube, Bus, and Bike.
- Edges: Two types of connections are modeled: transit network edges (e.g., tube lines) and road network edges (categorized by type: primary, secondary, residential, etc.). Connections between different node types (e.g., walking between a bus stop and a tube station) are modeled as shortest paths on the road network.
- Labels: Land use intensity is defined by six Point of Interest (POI) categories (Office, Sustenance, Transport, Retail, Leisure, Residence) within a 1km catchment area around each node.
Heterogeneous Graph Neural Network (HGT) Model:
The core model is a Heterogeneous Graph Transformer designed to handle multiple node and edge types.
- Feature Transformation: Node features are projected into a unified space based on their specific node types.
- Multi-Head Attention: A multi-head attention mechanism computes message representations from source to target nodes, accounting for the specific types of source nodes, target nodes, and connecting edges.
- Aggregation: Information is aggregated from neighbors using attention weights to update node representations, followed by a fully connected layer to predict land use intensities.
Post-hoc Explainability:
To address the black-box issue, the framework employs two XAI methods:
- Feature Attribution: Using Integrated Gradients (and Gradient $\cdot$ Input) to quantify the importance of input features (e.g., passenger flow at specific times) on the predicted land use.
- Counterfactual Explanations (CEs): A novel notion defined for heterogeneous graphs to identify the minimal changes required in a subgraph (node features, node types, edge types, or structure) to shift a prediction from its current state to a desirable "mixed" land use state. This involves calculating a local subgraph dissimilarity metric that accounts for feature, type, and structural differences.

Key Contributions

Data Integration: The authors construct a heterogeneous graph integrating multi-modal mobility data (Bus, Tube, Bike) and topology network data, enabling the learning of spatial representations that capture the heterogeneity of urban transport systems.
Novel Application of HGNs: This work pioneers the application of Heterogeneous Graph Networks for land use inference, specifically addressing spatial dependencies and the heterogeneity of nodes and edges. The study demonstrates that HGNs significantly outperform baseline models (NN, GCN, GAT, GraphSage, RGCN).
Explainability Framework: The paper introduces feature attribution analysis to verify model behavior against stakeholder expectations (e.g., commuter patterns) and defines a new counterfactual explanation method tailored for heterogeneous graphs. This allows planners to understand not just what the model predicts, but how specific input modifications could lead to different, ideal outcomes.

Results
Empirical experiments were conducted on a dataset covering inner and central London, utilizing data from Transport for London (TfL) and OpenStreetMap.

Performance: The proposed HGT model consistently outperformed all baseline models across all six land use indicators. The most significant improvements were observed in 'Sustenance' (68.35% improvement in MAE over NN) and 'Office' (63.95% improvement).
Spatial Analysis: Residual distribution analysis showed that HGT significantly reduced the systematic overestimation or underestimation of specific areas compared to homogeneous models (GCN), indicating better handling of spatial heterogeneity.
Ablation Study: Experiments removing specific transport modes confirmed that increased connectivity and heterogeneous information (combining bus, tube, and bike data) are complementary and crucial for prediction accuracy, particularly for challenging categories like 'Leisure'.
Explainability Findings:
- Feature Attribution: The importance of time-of-day features aligned with known human activity patterns (e.g., high morning inflow for residential areas, high evening inflow for office/sustenance areas).
- Counterfactuals: Analysis revealed that achieving a "mixed" land use state requires varying degrees of change depending on the location. For some nodes, changing node features (traffic flow) was the primary driver, while for others, changing edge types (road hierarchy) or node types was more critical.

Significance and Claims
The paper claims that the proposed framework provides a robust, accurate, and transparent tool for urban stakeholders. By moving beyond independent sample processing to a heterogeneous graph approach, the model captures complex spatial dependencies and service heterogeneity. Crucially, the integration of post-hoc explanations bridges the gap between high-accuracy deep learning and the transparency required for policy-making. The authors posit that these insights—specifically the ability to verify model logic against expert knowledge and to simulate "what-if" scenarios for land use planning—build trust in AI-driven urban planning tools. The work is modest in its scope, acknowledging limitations such as the use of 1-hop relationships and the reliance on outflow data, suggesting future work should explore multi-step transfers and more comprehensive data inputs.