config.yaml

# Configuration for Oxford Nanopore Polishing Decision Tool
# ==========================================================

# Data paths
data:
  input_csv: "data/all_samples_polishing_metrics.csv"
  merged_csv: "data/all_samples_polishing_metrics.csv"
  labeled_csv: "data/training_dataset_with_target.csv"
  engineered_csv: "data/training_dataset_engineered.csv"

# Class Configuration (CRITICAL IMPROVEMENT: 3-class system)
classes:
  n_classes: 3
  class_mapping:
    1: "Early (R1-R2)"  # Stop early: rounds 1-2
    2: "Medium (R3-R4)"  # Stop mid-way: rounds 3-4
    3: "Late (R5)"      # Continue to end: round 5
  
  # Original to new class mapping
  original_to_new:
    1: 1  # Early
    2: 1  # Early
    3: 2  # Medium
    4: 2  # Medium
    5: 3  # Late

# R1 Quality Thresholds (for early stopping consideration)
r1_thresholds:
  min_busco: 95.0           # Minimum BUSCO complete %
  max_assembly_error: 0.02  # Maximum |assembly_frac - 1.0|
  max_error_rate: 0.07      # Maximum error rate
  max_contigs: null         # Maximum contigs (null = ignore)
  min_coverage_est: 8.0     # Minimum coverage estimate
  max_align_err_cons: 0.15  # Maximum alignment error in consensus

# Stability Thresholds (for R1 stability check)
stability:
  eps_qv: 0.05              # QV tolerance
  eps_error: 0.0005         # Error rate tolerance
  eps_busco: 1.0            # BUSCO tolerance
  eps_assembly_frac: 0.01   # Assembly fraction tolerance
  use_assembly_frac: false  # Whether to include assembly_frac in stability

# Plateau Detection
plateau:
  relative_threshold: 0.12  # 12% of max improvement

# Feature Engineering
features:
  # Base features from data
  base:
    - n50
    - qv
    - error_rate
    - busco_complete
    - busco_fragmented
    - busco_missing
    - assembly_frac
    - assembly_error
    - num_contigs
    - total_length
  
  # Ratio features (NEW)
  ratios:
    - qv_per_round
    - busco_per_round
    - improvement_rate
    - cost_benefit_ratio
  
  # Trend features (NEW)
  trends:
    - delta_qv
    - delta_busco
    - delta_error
    - delta_assembly
    - score_improvement
  
  # Cumulative features
  cumulative:
    - gain_cum
    - plateau_streak
  
  # Normalized to R1 (NEW)
  normalized_r1:
    - qv_from_r1
    - busco_complete_from_r1
    - error_rate_from_r1
    - assembly_frac_from_r1

# Model Training Configuration
models:
  random_state: 42
  test_size: 0.20
  cv_folds: 5
  
  # XGBoost Configuration
  xgboost:
    n_estimators: 500
    max_depth: 6
    learning_rate: 0.05
    subsample: 0.8
    colsample_bytree: 0.8
    min_child_weight: 3
    gamma: 0.1
    reg_alpha: 0.1
    reg_lambda: 1.0
  
  # Random Forest Configuration
  random_forest:
    n_estimators: 800
    max_depth: 12
    min_samples_leaf: 2
    max_features: "sqrt"
  
  # Ordinal Regression Configuration
  ordinal_regression:
    alpha: 1.0

# Imbalance Handling
imbalance:
  use_smote: true
  smote_k_neighbors: 3
  smote_sampling_strategy: "auto"
  
  # Class weights
  class_weights:
    1: 2.0   # Early - boost minority
    2: 1.5   # Medium
    3: 1.0   # Late - most common

# Evaluation Metrics
evaluation:
  primary_metric: "balanced_accuracy"
  target_metrics:
    balanced_accuracy: 0.65
    macro_f1: 0.60
    min_class_recall: 0.40
    mae: 0.5
    qwk: 0.50

# Hierarchical Policy (3-class version)
hierarchical:
  stage_a:
    # Binary: Early (1) vs Continue (2-3)
    threshold: 0.5
    model: "logistic_regression"
  
  stage_b:
    # Binary: Medium (2) vs Late (3)
    model: "random_forest"

# Post-processing Rules
postprocessing:
  conservative_bias: true  # Prefer one extra round over stopping too early
  smooth_predictions: true # Avoid large jumps in predictions
  use_domain_rules: true   # Apply r1_ok and stability rules

# Output Configuration
outputs:
  models_dir: "models/"
  results_dir: "outputs/"
  plots_dir: "outputs/plots/"
  logs_dir: "logs/"

# Logging
logging:
  level: "INFO"
  format: "%(asctime)s - %(name)s - %(levelname)s - %(message)s"
  file: "logs/pipeline.log"