LEGN-CARA

# LEGN-CARA: Graph-based Recommendation System for Respiratory Health Risk Mitigation ## 🌍 Overview This repository presents **LEGN-CARA**, a novel computational framework that leverages **graph neural networks (GNNs)** and **counterfactual analysis** to mitigate respiratory health risks associated with **climate change** and **biodiversity loss**. The system integrates environmental exposure data, spatial-temporal modeling, and health outcomes to deliver **personalized and interpretable health risk assessments**. ## 📌 Key Features - **Latent Exposure Graph Network (LEGN)**: Models complex spatial and temporal relationships between environmental exposures.- **Counterfactual-Augmented Risk Attribution (CARA)**: Simulates hypothetical exposure scenarios to assess individual and population-level risk attributions.- **Multi-modal Data Fusion**: Integrates environmental, behavioral, and health data streams.- **Interpretable AI**: Gradient-based attribution and attention mechanisms provide transparent insights into risk factors.- **Scalable Deployment**: Modular design suitable for low-resource settings and diverse geographies. ## 🧠 Methodology ### 1. Latent Exposure Graph Network (LEGN)- Uses **graph-based representation learning** to capture interactions between pollutants and environmental features.- Applies **temporal encoding** using GRUs and **self-attention** for long-range dependencies.- Includes a **spatial fusion module** to incorporate neighborhood-level information. ### 2. Counterfactual-Augmented Risk Attribution (CARA)- Computes risk scores by simulating changes in exposure values.- Introduces **stability and sparsity constraints** for robust attribution.- Enhances interpretability via **information-theoretic calibration**. ## 🧪 Experiments Experiments were conducted on standard human pose datasets (COCO, MPII, Human3.6M, LSP), treating keypoint detection tasks as proxy evaluations for model adaptability and robustness. ### 📈 Performance Highlights:| Dataset | Accuracy | Recall | F1 Score | AUC ||---------|----------|--------|----------|-----|| COCO | 91.35% | 88.43% | 89.76% | 92.28% || MPII | 90.21% | 86.55% | 88.18% | 91.03% || Human3.6M | 90.62% | 87.84% | 88.93% | 91.76% || LSP | 89.35% | 85.66% | 87.41% | 90.08% | ### ⚙️ Ablation Studies:Confirmed the necessity of:- LEGN (graph modeling)- CARA (counterfactual reasoning)- Temporal self-attention ## 📂 Structure