Causal Discovery Algorithm Selection

Causal Discovery Algorithm Selection Meta-Learner A meta-learning system that predicts the for any discrete observational dataset, based on dataset meta-features. ## 🎯 What it Does Given a new discrete dataset (pandas DataFrame), the system: 1. (entropy, mutual information, chi2 statistics, CI test probes, etc.) 2. for each of 9 algorithms via trained models 3. algorithms expected to produce the most accurate CPDAG ## 📊 Performance (Leave-One-Network-Out Cross-Validation) ### Best Model: Pairwise-GBM Ranking | Metric | Value | |--------|-------| | | (true best algorithm is in predicted top-3) | | | (tiny SHD gap vs oracle selection) | | | (majority of predictions are perfect) | ### Model Comparison (178 configs, 14 networks + augmented) | Model | Top-3 Hit Rate | NDCG@3 | Mean Regret | |-------|---------------|--------|-------------| | | | — | 0.011 | | GBM-300-lr01 | 67.4% | 0.957 | 0.011 | | RF-200 | 66.9% | 0.961 | 0.007 | | RF-500 | 66.3% | 0.962 | 0.007 | | GBM-500-lr05 | 65.2% | 0.948 | 0.013 | ### Progression | Stage | Configs | Networks | Top-3 Hit Rate | |-------|---------|----------|---------------| | Initial (small nets) | 65 | 4 | 68.2% | | All 14 networks | 122 | 14 | 70.5% | | + Data augmentation | 178 | 14+aug | | ## 🧪 Algorithm Pool (9 algorithms) | Algorithm | Family | Library | Output | Wins | |-----------|--------|---------|--------|------| | | Score-based | causal-learn | CPDAG | 47% | | | Constraint-based | causal-learn | CPDAG | 32% | | | Constraint-based | causal-learn | PAG | 8% | | | Score-based | pgmpy | DAG | 6% | | | Score-based (greedy) | pgmpy | DAG | 3% | | | Score-based (meta) | pgmpy | DAG | 2% | | | Permutation-based | causal-learn | CPDAG | 1% | | | Permutation-based | causal-learn | CPDAG | 1% | | | Hybrid | pgmpy | DAG | =0.1.4 pgmpy>=0.1.25 scikit-learn>=1.8 pandas numpy scipy joblib ``` ## 📚 References - (arxiv:2504.13263) — Closest existing algorithm selection system - (arxiv:2205.12934) — Amortized causal structure learning (biaffine architecture) - (arxiv:2602.08629) — Scalable neural causal discovery - (arxiv:1611.01734) — Deep Biaffine Attention for dependency parsing - (arxiv:2005.00975) — Global structural training loss for parsing - (arxiv:1401.2474) — Algorithm selection via meta-learning - (bnlearn.com) — Bayesian network benchmark repository ## 🔮 Future Work (Phase 2) 1. : Pre-train a neural feature extractor that learns variable-pair "arc scores" 2. (TreeCRF-inspired): Global ranking optimization instead of per-algorithm MSE 3. : Predict not just which algorithm but optimal hyperparameters (CASH) 4. : Run top-3 and vote on edges across their CPDAGs ## License MIT