arminmehrabian/nasa-eosdis-heterogeneous-gnn

EOSDIS Graph Neural Network Model Card

Model Overview

Model Name: EOSDIS-GNN Version: 1.0.4 Type: Heterogeneous Graph Neural Network Framework: PyTorch + PyTorch Geometric Base Language Model: nasa-impact/nasa-smd-ibm-st-v2

This model was trained on 2025-09-10.

Model Details

• Hidden Channels: 256
• Number of Layers: 3
• Convolution Type: sage
• Max Epochs: 100000

Core Components

• Base Text Encoder: NASA-SMD-IBM Language Model (768-dimensional embeddings)
• Graph Neural Network: Heterogeneous GNN with multiple layers
• Node Types: Dataset, Publication, Instrument, Platform, ScienceKeyword
• Edge Types: Multiple relationship types between nodes

Technical Specifications

• Input Dimensions: 768 (NASA-SMD-IBM embeddings)
• Hidden Dimensions: Configurable (default: 256)
• Output Dimensions: 768 (aligned with NASA-SMD-IBM space)
• Number of Layers: Configurable (default: 3)
• Activation Function: ReLU
• Dropout: Applied between layers

Training Details

Training Data

• Source: NASA EOSDIS Knowledge Graph
• Node Types and Counts:
  • Datasets: Earth science datasets from NASA DAACs
  • Publications: Related scientific papers
  • Instruments: Earth observation instruments
  • Platforms: Satellite and other observation platforms
  • Science Keywords: NASA Earth Science taxonomy

Training Process

• Optimization: Adam optimizer
• Loss Function: Contrastive loss for semantic alignment
• Training Strategy:
  • Initial node embedding generation
  • Message passing through graph structure
  • Contrastive learning with NASA-SMD-IBM embeddings

Intended Use

Designed for: research, data discovery, and semantic search in Earth science
Not intended for: safety‑critical systems or unrelated domains without fine‑tuning

Strengths

1. Semantic Understanding:

   • Strong performance in finding semantically related content
   • Effective cross-modal relationships between text and graph structure

2. Domain Specificity:

   • Specialized for Earth science terminology
   • Understands relationships between instruments, platforms, and datasets

3. Multi-modal Integration:

   • Combines text-based and graph-based features
   • Preserves domain-specific relationships

Limitations

1. Data Coverage:

   • Performance depends on training data coverage
   • May have gaps in newer or less documented areas

2. Computational Requirements:

   • Requires significant memory for full graph processing
   • Graph operations can be computationally intensive

3. Domain Constraints:

   • Optimized for Earth science domain
   • May not generalize well to other domains

Usage Guide

Installation Requirements

pip install torch torch-geometric transformers huggingface-hub

Basic Usage

from transformers import AutoTokenizer, AutoModel
import torch
from gnn_model import EOSDIS_GNN

# Load models
tokenizer = AutoTokenizer.from_pretrained("nasa-impact/nasa-smd-ibm-st-v2")
text_model = AutoModel.from_pretrained("nasa-impact/nasa-smd-ibm-st-v2")
gnn_model = EOSDIS_GNN.from_pretrained("your-username/eosdis-gnn")

# Process query
def get_embedding(text):
    inputs = tokenizer(text, return_tensors="pt", max_length=512, 
                      truncation=True, padding=True)
    with torch.no_grad():
        outputs = text_model(**inputs)
        return outputs.last_hidden_state[:, 0, :]

Semantic Search Example

from semantic_search import SemanticSearch

# Initialize searcher
searcher = SemanticSearch()

# Perform search
results = searcher.search(
    query="atmospheric carbon dioxide measurements",
    top_k=5,
    node_type="Dataset"  # Optional: filter by node type
)

Evaluation Metrics

---

Performance

Metric	Value	Notes
Top‑5 Accuracy	87.4%	Probability that at least one of the top‑5 retrieved nodes is relevant.
Mean Reciprocal Rank (MRR)	0.73	Measures ranking quality.
Link Prediction ROC‑AUC	0.91	Ability to predict whether a given edge exists.
Node Classification F1 (macro)	0.84	Balanced accuracy across node types.
Triple Classification Accuracy	88.6%	Accuracy in classifying valid vs. invalid triples.

Evaluation Notes:

• Dataset: held‑out portion of NASA EOSDIS Knowledge Graph
• Search task: queries derived from publication abstracts
• Link prediction: 80/10/10 train/val/test splits
• Numbers from offline evaluation; may vary on different graph snapshots

Version Control

• Model versions tracked on Hugging Face Hub
• Regular updates for improved performance

Citation

@misc{armin_mehrabian_2025,
    author       = { Armin Mehrabian },
    title        = { nasa-eosdis-heterogeneous-gnn (Revision 7e71e62) },
    year         = 2025,
    url          = { https://huggingface.co/arminmehrabian/nasa-eosdis-heterogeneous-gnn },
    doi          = { 10.57967/hf/6071 },
    publisher    = { Hugging Face }
}

Contact Information

• Maintainer: Armin Mehrabian
• Email: armin.mehrabian@nasa.gov
• Organization: NASA