Skip to content

Tech Specs

Default Coarse Search Parameter Grids (by Mode)

Model Parameter fast Mode industrial Mode
Logistic clf__C [0.01, 0.1, 1, 10] [0.01, 0.1, 1, 10]
clf__penalty ['l1', 'l2'] ['l1', 'l2', 'elasticnet']
clf__l1_ratio Not used [0.1, 0.5, 0.9]
RandomForest clf__n_estimators [100, 200] [300, 500, 1000]
clf__max_depth [5, 10] [10, 20, None]
clf__min_samples_split [2, 5, 10] [2, 5, 10]
XGBoost clf__n_estimators [100, 200] [300, 500, 1000]
clf__max_depth [3, 5] [3, 5, 7]
clf__learning_rate [0.1, 0.2] [0.01, 0.05, 0.1]
clf__subsample Not used [0.6, 0.8, 1.0]
LightGBM clf__n_estimators [100, 200] [300, 500, 1000]
clf__max_depth [5, 10] [10, 15, 20]
clf__learning_rate [0.1, 0.2] [0.01, 0.05, 0.1]
clf__num_leaves Not used [15, 31, 63]
CatBoost clf__iterations [200, 300] [500, 1000]
clf__depth [4, 6] [4, 6, 8]
clf__learning_rate [0.1, 0.2] [0.01, 0.05, 0.1]
clf__l2_leaf_reg Not used [1, 3, 5, 7]

✅ fast → lightweight grid for quick experimentation
✅ industrial → larger, production-ready grid covering more parameter combinations

You can set this using the mode argument when calling build_model_config() or initializing GridMaster.

Model Random State Special Defaults Recommended Use Cases
Logistic 66 solver='liblinear' if ≤10,000 samples or mode='fast'; solver='saga' with penalty=['l1', 'l2', 'elasticnet'] if >10,000 samples or mode='industrial' Best for small-to-medium datasets or when you need interpretable models; saga supports large-scale data and elasticnet but requires standardized inputs.
RandomForest 66 Uses sklearn RandomForestClassifier; adjusts trees and depth based on mode Excellent general-purpose model, robust to overfitting, works well on tabular data with mixed feature types; fast mode for quick trials, industrial mode for robust tuning.
XGBoost 66 eval_metric='logloss', use_label_encoder=False, verbosity=0; optional GPU configs allowed via custom_estimator_params Highly performant on structured data, handles missing values natively; recommended for competition-grade or production tasks; supports GPU for large-scale runs.
LightGBM 66 verbosity=-1; optional GPU configs allowed via custom_estimator_params Similar to XGBoost but faster on large datasets; works well with categorical features; recommended for fast iteration and industrial pipelines.
CatBoost 66 verbose=0; optional GPU configs allowed via custom_estimator_params Best when working with categorical data; often requires less parameter tuning out-of-the-box; GPU acceleration improves scalability on big data.

Fine Search Modes: Parameter Details Table

Model Smart Mode (Auto Top 2) Expert Mode (Pre-Selected) Custom Mode (User-Defined)
Logistic Regression Based on top test score variation — usually clf__C, clf__penalty Always clf__C Whatever user provides in custom_fine_params
Random Forest Based on top test score variation — usually clf__max_depth, clf__min_samples_split Always clf__max_depth, clf__min_samples_split Whatever user provides in custom_fine_params
XGBoost Based on top test score variation — usually clf__learning_rate, clf__max_depth Always clf__learning_rate, clf__max_depth Whatever user provides in custom_fine_params
LightGBM Based on top test score variation — usually clf__learning_rate, clf__max_depth Always clf__learning_rate, clf__max_depth Whatever user provides in custom_fine_params
CatBoost Based on top test score variation — usually clf__learning_rate, clf__depth Always clf__learning_rate, clf__depth Whatever user provides in custom_fine_params

Mode Selection Key Points

  • Smart Mode dynamically detects which parameters matter most for each dataset — great for flexible, adaptive tuning.
  • Expert Mode sticks to proven influential parameters, reducing grid size and focusing search.
  • Custom Mode gives you complete freedom but requires you to define a meaningful and valid parameter grid yourself.

Parallelization Strategy

By default, GridMaster uses half of the detected CPU cores (n_jobs) to balance system load and optimization speed.

You can override this by setting:

  • n_jobs = -1 → use all available cores.

  • n_jobs = <int> → explicitly set the number of parallel jobs.

Tip: On shared or production servers, test carefully before using full CPU to avoid resource contention.

GPU Acceleration Support

Supported for:

  • XGBoost → via tree_method='gpu_hist'

  • LightGBM → via device='gpu'

  • CatBoost → via task_type='GPU'

These can be passed through the custom_estimator_params argument.

Important: Requires proper GPU drivers and library installations; otherwise, the system may silently fall back to CPU without warnings.

Pipeline Preprocessing Details

Logistic Regression:

  • Always uses a StandardScaler for feature normalization.

Tree-based models (Random Forest, XGBoost, LightGBM, CatBoost):

  • Use 'passthrough' because they are scale-invariant and don’t require normalization.

Evaluation Metric Defaults

Logistic Regression, Random Forest:

  • Follow sklearn’s default scoring (accuracy).

XGBoost:

  • Uses eval_metric='logloss' by default.

LightGBM, CatBoost:

  • Rely on their own internal defaults unless overridden.