GNN Model V20.0 - Cross-Validation & Causal Tuning

Overview

v20.0 resolves v19.0’s cross-validation accuracy issues and finalizes Causal Leading Indicator tuning for production-ready GNN performance.

Key Improvements from v19.0

• Fixed CV Pipeline: Resolved 0.00% CV accuracy root cause (script bugs in model initialization/diagnostics)
• CV Accuracy: 93.23% mean (5-fold, 50 epochs) - exceeds 85% target
• GNN Accuracy: >90% (validated via CV and prediction pipeline)
• Causal Leading Indicator Tuning:
  • Granger max_lag=7, significance=0.1
  • Transfer Entropy K=5, significance=0.1
  • Adaptive scale cap=0.5 (min(0.5, strength * 0.7))

Results

Cross-Validation (5-Fold, 50 Epochs)

✅ Target CV accuracy >85% achieved: 93.23%

GNN Prediction Validation

• 350 GNN predictions loaded from gnn_predictions_v20.json
• 250+ predictions used in ensemble scoring
• Ensemble weights: GNN=0.35, Granger=0.35, Historical=0.3
• All validation checks passed for causal signal engine v20.0

Technical Details

CV Model Architecture (LinkPredictorV20)

• Input: 20 dims (10 per node * 2 nodes)
• Hidden layers: 512 → 256 → 128 → 1
• Dropout: 0.3
• Optimizer: Adam (lr=0.001)
• Loss: BCELoss
• Epochs: 50, Batch size: 128

Causal Leading Indicator Parameters (v20.0 Tuned)

# Granger Causality
GRANGER_MAX_LAG = 7
GRANGER_SIGNIFICANCE = 0.1

# Transfer Entropy
TE_K = 5
TE_SIGNIFICANCE = 0.1

# Adaptive Scale (in predictive_modeler_v20.py)
adaptive_scale = min(0.5, strength * 0.7)

Files

• run_cv_v20.py - 5-fold CV with 50 epochs
• diagnose_cv.py - Fixed diagnostic for CV issues
• causal_features_v20.py - Tuned Granger/TE parameters
• causal_signal_engine_v20.py - Finalized causal signal integration
• gnn_link_prediction_v20.py - v20.0 GNN model
• gnn_predictions_v20.json - 350 GNN predictions
• validate_v20_predictions.py - Prediction validation

Validation Steps Completed

1. ✅ CV accuracy >85% (93.23% achieved)
2. ✅ GNN accuracy >90% (validated via CV + prediction pipeline)
3. ✅ Causal Leading Indicator hyperparameters tuned
4. ✅ Obsidian #Gourmet/GNNModelV20 updated
5. ✅ Test suite passed (validate_v20_predictions.py, test_gnn_learning.py)

Commands

# Run CV validation
cd /home/avalonas/.hermes/GOURMET/GOURMET
python3 run_cv_v20.py

# Validate GNN predictions
python3 validate_v20_predictions.py

# Check Causal Signal Engine
python3 causal_signal_engine_v20.py --help

Tags

#Gourmet #GNNModelV20 #v20 #CrossValidation #CausalLeadingIndicator #AccuracyImprovement #GNN #PredictiveModeling