config_optimization.yaml

# MALLORN TDE Classification Optimization Configuration
# Target: F1 = 0.76 (from baseline 0.66)

# =============================================================================
# GENERAL SETTINGS
# =============================================================================
seed: 42
n_folds: 5
n_repeats: 2
output_dir: "./submissions"

# =============================================================================
# FEATURE ENGINEERING
# =============================================================================
features:
  # Core features
  use_bayesian_blocks: true
  use_temporal_morphology: true
  use_color_evolution: true
  
  # FATS (Feature Analysis for Time Series)
  use_fats_features: true
  fats_exclude_percentile_ratios: true  # Exclude problematic features
  
  # Additional physics features
  use_power_law_fit: true
  use_variability_indices: true
  use_cross_band_features: true

# =============================================================================
# SEMI-SUPERVISED LEARNING
# =============================================================================
semi_supervised:
  enabled: true
  method: "kmeans"
  n_clusters: 40           # K in [35, 50] range
  confidence_threshold: 0.80  # Min purity for pseudo-labels
  outlier_percentile: 95      # Exclude top 5% outliers
  
  # Alternative: Confidence-based pseudo-labeling
  use_confidence_threshold: true
  high_confidence_threshold: 0.95  # For TDE class
  medium_confidence_threshold: 0.80  # For SN/AGN

# =============================================================================
# DATA AUGMENTATION
# =============================================================================
augmentation:
  # SMOTE configuration
  use_smote: true
  k_neighbors: 3           # Reduced for small TDE sample (64)
  sampling_strategy: 0.1   # Conservative 1:10 ratio
  
  # Alternative SMOTE variants
  use_borderline_smote: false
  use_svm_smote: false
  use_smote_enn: false
  
  # Physics-based augmentation
  use_temporal_stretch: true
  stretch_range: [0.8, 1.2]
  use_photometric_noise: true
  noise_scale: 0.05

# =============================================================================
# MODEL CONFIGURATION
# =============================================================================
models:
  # LightGBM (Primary model)
  lightgbm:
    enabled: true
    num_leaves: 31
    max_depth: 6
    learning_rate: 0.05
    n_estimators: 2000
    min_child_samples: 20
    feature_fraction: 0.8
    bagging_fraction: 0.8
    bagging_freq: 5
    
  # XGBoost
  xgboost:
    enabled: true
    max_depth: 6
    learning_rate: 0.05
    n_estimators: 2000
    subsample: 0.8
    colsample_bytree: 0.8
    min_child_weight: 5
    
  # Random Forest
  random_forest:
    enabled: true
    n_estimators: 500
    max_depth: 12
    min_samples_split: 10
    min_samples_leaf: 4
    class_weight: "balanced"
    
  # Extra Trees
  extra_trees:
    enabled: true
    n_estimators: 500
    max_depth: 12
    min_samples_split: 10
    min_samples_leaf: 4
    class_weight: "balanced"

# =============================================================================
# CLASS IMBALANCE HANDLING
# =============================================================================
class_imbalance:
  # Class weights (inverse frequency)
  # TDE:SN:AGN = 64:727:1407 → weights ≈ 22:2:1
  class_weights:
    0: 22.0   # TDE (minority)
    1: 2.0    # SN
    2: 1.0    # AGN (majority)
  
  # Alternative weighting schemes
  use_sqrt_balanced: false  # Use 1/sqrt(n) instead of 1/n
  
  # Cost-sensitive learning
  use_cost_sensitive: true
  misclassification_costs:
    tde_fn: 22.0  # Penalty for missing TDE
    sn_fn: 2.0
    agn_fn: 1.0

# =============================================================================
# THRESHOLD OPTIMIZATION
# =============================================================================
threshold:
  enabled: true
  method: "pr_curve"  # "pr_curve" or "cost_based"
  
  # Precision-Recall curve optimization
  search_range: [0.1, 0.9]
  search_step: 0.01
  
  # Class-specific thresholds
  # TDE typically needs lower threshold (0.2-0.4)
  default_threshold: 0.5
  tde_threshold: 0.25
  sn_threshold: 0.5
  agn_threshold: 0.75
  
  # Cost-based optimization
  optimize_for: "f1_macro"  # "f1_macro", "f1_weighted", "cost"

# =============================================================================
# ENSEMBLE CONFIGURATION
# =============================================================================
ensemble:
  method: "average"  # "average", "weighted", "stacking"
  
  # Model weights for weighted ensemble
  weights:
    lightgbm: 0.35
    xgboost: 0.30
    random_forest: 0.20
    extra_trees: 0.15
  
  # Stacking meta-learner
  use_stacking: false
  meta_learner: "logistic_regression"

# =============================================================================
# VALIDATION
# =============================================================================
validation:
  method: "stratified_kfold"
  n_splits: 5
  shuffle: true
  
  # Check for data leakage
  check_leakage: true
  group_column: "object_id"  # Ensure same object not in train/val
  
  # Stratification
  maintain_class_ratio: true
  min_samples_per_class: 3   # Minimum samples per class per fold

# =============================================================================
# ADVANCED OPTIONS
# =============================================================================
advanced:
  # Probability calibration
  calibrate_probabilities: true
  calibration_method: "isotonic"  # "sigmoid" (Platt) or "isotonic"
  
  # Feature selection
  use_feature_selection: true
  n_features_select: 300
  selection_method: "mutual_info"  # "mutual_info", "chi2", "f_classif"
  
  # Dimensionality reduction
  use_pca: false
  pca_variance: 0.95
  
  # Label smoothing
  use_label_smoothing: false
  smoothing_alpha: 0.1

# =============================================================================
# EXPECTED PERFORMANCE BY CONFIGURATION
# =============================================================================
performance_targets:
  baseline: 0.66
  
  # With physics features only
  phase1_physics_features: 0.70
  
  # With threshold optimization
  phase2_threshold_opt: 0.73
  
  # With semi-supervised learning
  phase3_semi_supervised: 0.76
  
  # Full optimization
  target: 0.76

# =============================================================================
# TROUBLESHOOTING GUIDE
# =============================================================================
troubleshooting:
  # If F1 < 0.70
  low_f1:
    - "Check Bayesian block features are enabled"
    - "Verify FATS features exclude percentile ratios"
    - "Ensure class weights are applied"
    - "Check for data leakage"
    
  # If overfitting
  overfitting:
    - "Reduce max_depth to 5-6"
    - "Increase min_child_samples to 30"
    - "Reduce num_leaves to 20"
    - "Increase bagging_fraction"
    
  # If poor TDE recall
  poor_recall:
    - "Lower TDE threshold to 0.2"
    - "Increase TDE class weight to 30"
    - "Add more SMOTE samples"
    - "Use focal loss for hard examples"
    
  # If unstable CV
  unstable_cv:
    - "Increase n_folds to 10"
    - "Use stratified shuffle split"
    - "Set larger random seed range"
    - "Check for duplicate objects"