Literature Overview - davidlabee/Graph4Air GitHub Wiki
📚 Literature Overview
Overview
This section serves as a curated and annotated bibliography of the most relevant academic work related to our project. It includes foundational work on Graph Neural Networks (GNNs), graph construction strategies, and applications of GNNs in spatial or environmental data modeling.
📖 1. GNN Fundamentals
Purpose: To understand the core architectures and ideas behind Graph Convolutional Networks, Graph Attention Networks, and other foundational models.
📍A Gentle Introduction to Graph Neural Networks
This article introduces modern graph neural networks by explaining graph-structured data, highlighting the unique challenges of graph modeling, guiding readers through the development of a GNN from basics to advanced models, and offering a hands-on playground to deepen understanding.
🏙 2. GNNs for Spatial or Environmental Applications
Purpose: To find prior work where GNNs are used for tasks like pollution modeling, traffic prediction, or spatial interpolation.
They developed a graph-based machine learning framework to estimate pedestrian-specific PM2.5 exposure in New York City by combining vehicle-sensed pollution data with high-resolution pedestrian activity estimates, revealing overlooked high-risk zones and offering new insights for pedestrian-centered urban air quality management.
They developed a GNN model to quickly predict traffic volume changes from policy interventions—like capacity reductions—using simulated data from Paris, offering a faster alternative to traditional traffic simulations, especially effective on major roads.
📍Uncertainty-aware probabilistic graph neural networks for road-level traffic crash prediction
They proposed STZITD-GNN, a novel uncertainty-aware graph neural network model that combines statistical Tweedie distributions with GNNs to predict road-level daily traffic crash risks—including high, low, and zero-risk cases—more accurately and reliably than existing models.
They developed a deep learning model combining Graph Neural Networks (GNNs) and Gated Recurrent Units (GRUs) to predict 48-hour PM2.5 levels in Taiwan, including in areas without monitoring stations.
📍Short-term air pollution prediction using graph convolutional neural networks
They developed and evaluated spatio-temporal graph convolutional models (STGCN-A/B/C) for short-term air pollution forecasting, showing that their models—especially STGCN-C—outperform baselines in predicting PM2.5 levels in Delhi and California using multivariate data across different time horizons.
📍Opportunistic Air Quality Monitoring and Forecasting with Expandable Graph Neural Networks
They proposed an Expandable Graph Attention Network (EGAT) model that adapts to evolving and flexible air quality monitoring infrastructures, enabling accurate forecasting even in areas with newly added or sparse data sources.
They developed a spatio-temporal GNN-based model to forecast hourly PM2.5 levels in California and integrated it with prescribed fire simulations to assess air quality impacts, helping identify optimal timing—like March—for conducting prescribed fires.
📍Graph Neural Network for Air Quality Prediction: A Case Study in Madrid
They proposed an Attention Temporal Graph Convolutional Network (combining Attention, GRU, and GCN) for air quality prediction, which outperformed baseline models on Madrid's air, weather, and traffic data in terms of accuracy and correlation metrics.
They proposed a Context-augmented Graph Autoencoder (Con-GAE) model that detects anomalies in air quality data by constructing spatiotemporal graph structures from multivariate and multi-station data, effectively capturing spatial and temporal correlations for improved abnormal event detection.
📍Interpretable Crowd Flow Prediction with Spatial-Temporal Self-Attention
They proposed a Spatial-Temporal Self-Attention Network (STSAN) that avoids splitting spatial and temporal features by using a unified encoding gate and positional/time encodings. Their model uses a multi-aspect attention mechanism to directly capture and interpret complex spatial-temporal dependencies. It outperformed state-of-the-art models by reducing inflow and outflow RMSE by 16% and 8% respectively on the Taxi-NYC dataset.
📍 SOURCES ON AIR QUALITY WITH MOBILE SENSORS?
🧩 3. Graph Construction Techniques
Purpose: To explore strategies for building graphs from structured/unstructured data (e.g., k-NN, thresholded distances, spatial proximity, semantic similarity).
📍 Graph construction on complex spatiotemporal data for enhancing machine learning performance
Performs an in-depth analysis of several methods for constructing graphs from a set of sensors attributed with spatial information, emphasizing the importance of appropriate graph construction in enhancing machine learning performance on spatiotemporal data. 
📍 Efficient K-Nearest Neighbor Graph Construction for Generic Similarity Measures
Presents NN-Descent, a simple yet efficient algorithm for approximate k-NN graph construction with arbitrary similarity measures, suitable for large-scale applications.
📍 Construction and Inference Method of Semantic-Driven, Spatio-Temporal Knowledge Graphs
Proposes a method that combines qualitative descriptions with quantitative distances to improve the accuracy of geospatial queries, enhancing the expressiveness and adaptability of traditional quantitative queries. 📍 Construction and Inference Method of Semantic-Driven, Spatio-Temporal Knowledge Graphs
📍Add literature describing how graphs are constructed from raw geospatial or road-level data.