MzeeChakulaNutritionEnsembleModel

MzeeChakula Nutrition Ensemble

Graph Neural Network Ensemble for Elderly Nutrition Planning in Uganda

Model Description

Production-ready ensemble combining three Graph Neural Networks:

• CRGN (40%) - Compositional reasoning
• HetGNN (35%) - Heterogeneous graph handling
• GAT (25%) - Attention mechanisms

Performance:

• 5.36M parameters (~20.4 MB)
• ~15ms inference (CPU), ~3ms (GPU)
• 90% link prediction AUC

Quick Start

from huggingface_hub import snapshot_download
import sys

# Download model
model_path = snapshot_download(repo_id="YOUR_USERNAME/mzeechakula-nutrition")
sys.path.append(model_path)

# Load and use
from nutrition_ensemble import NutritionEnsemble

ensemble = NutritionEnsemble(model_path)
recommendations = ensemble.recommend(user_id=0, top_k=10)
print(recommendations)

Installation

pip install torch numpy pandas huggingface_hub

Usage

Basic Recommendations

from nutrition_ensemble import NutritionEnsemble

# Initialize
ensemble = NutritionEnsemble(".")

# Get top 10 recommendations
recs = ensemble.recommend(user_id=0, top_k=10)
print(recs)

With Exclusions

# Exclude already consumed foods
recs = ensemble.recommend(
    user_id=0,
    top_k=10,
    exclude_foods=[1, 2, 3]  # Food IDs to exclude
)

Model Details

• Developed by: Nannyombi Shakiran, Yul Lam Gatkuoth
• Institution: Makerere University
• Supervisor: Ggaliwango Marvin
• License: MIT
• Model type: Graph Neural Network Ensemble
• Task: Food recommendation for elderly nutrition

Training Data

• 1,048 nodes (foods, nutrients, conditions, regions)
• 14,359 edges (relationships)
• Uganda-specific nutrition knowledge graph

Intended Use

Personalized nutrition recommendations for elderly populations in Uganda:

• Health conditions (hypertension, diabetes, anemia, frailty, etc.)
• Regional food availability (Central, Western, Eastern, Northern)
• Cultural preferences (English and Luganda support)
• Cost-conscious recommendations (~4,232 UGX per meal)

Limitations

• Trained specifically on Ugandan food data
• Requires user health profile for optimal recommendations
• Not a substitute for professional medical advice
• Best suited for elderly population (60+ years)

Performance Benchmarks

CPU (Intel i7):

• Single prediction: 0.5 ms
• Top-10 recommendations: 15 ms
• Batch (100 users): 1.5 seconds

GPU (NVIDIA T4):

• Single prediction: 0.1 ms
• Top-10 recommendations: 3 ms
• Batch (100 users): 0.3 seconds

Repository Structure

.
├── nutrition_ensemble.py      # Main model class
├── ensemble_config.json       # Model configuration
├── requirements.txt           # Dependencies
├── embeddings/                # Pre-computed embeddings
│   ├── crgn_embeddings.npy   # 7.30 MB
│   ├── hetgnn_embeddings.npy # 0.73 MB
│   └── gat_embeddings.npy    # 1.22 MB
└── examples/
    └── quickstart.py          # Usage examples

Citation

@software{mzeechakula2025,
  title={MzeeChakula: Graph-Enhanced Nutrition Planning for Elderly in Uganda},
  author={Nannyombi, Shakiran and Gatkuoth, Yul Lam},
  year={2025},
  institution={Makerere University},
  supervisor={Ggaliwango, Marvin},
  url={https://github.com/MzeeChakula}
}

Links

• GitHub: https://github.com/MzeeChakula/models
• Data Repository: https://github.com/MzeeChakula/data
• Issues: https://github.com/MzeeChakula/models/issues

Contact

• Email: [email protected]
• Collaboration: [email protected]

---

Made for Uganda's elderly population 🇺🇬