Preset Models¶

OG-Learn provides carefully tuned preset configurations for both HV and LV models.

High-Variance (HV) Presets¶

These models are used in Stage 1 for pseudo-label generation.

LightGBM Presets¶

lightgbm (default)biglightgbm

Standard LightGBM configuration for most use cases.

OGModel(hv='lightgbm', lv='mlp')

Parameter	Value
n_estimators	500
num_leaves	1200
max_depth	9
learning_rate	0.05

Larger LightGBM for complex datasets.

OGModel(hv='biglightgbm', lv='mlp')

Parameter	Value
n_estimators	500
num_leaves	1200
max_depth	9
min_child_samples	1

XGBoost¶

OGModel(hv='xgboost', lv='mlp')

Parameter	Value
n_estimators	500
max_depth	11
learning_rate	0.05

CatBoost¶

OGModel(hv='catboost', lv='mlp')

Parameter	Value
iterations	500
depth	9
learning_rate	0.05

Other HV Models¶

# Random Forest
OGModel(hv='random_forest', lv='mlp')

# Decision Tree
OGModel(hv='decision_tree', lv='mlp')

# Linear Regression (baseline)
OGModel(hv='linear_regression', lv='mlp')

Low-Variance (LV) Presets¶

These neural network models are used in Stage 2 for learning from pseudo-labels.

MLP (Multi-Layer Perceptron)¶

mlp (default)bigmlp

Standard MLP architecture.

OGModel(hv='lightgbm', lv='mlp')

Parameter	Value
hidden_layers	[256, 128, 64]
dropout	0.3
batch_size	256
learning_rate	0.001

Larger MLP for complex patterns.

OGModel(hv='lightgbm', lv='bigmlp')

Parameter	Value
hidden_layers	[512, 256, 128]
dropout	0.3
batch_size	256
learning_rate	0.001

ResNet¶

Deep residual network with skip connections.

OGModel(hv='lightgbm', lv='resnet')

Parameter	Value
d_main	256
d_hidden	512
n_blocks	2
dropout	0.2

Transformer¶

Attention-based architecture for capturing complex dependencies.

OGModel(hv='lightgbm', lv='transformer')

Parameter	Value
n_layers	3
d_token	192
n_heads	8
d_ffn	256

Listing Available Presets¶

from og_learn import list_presets

list_presets()

Output:

============================================================
              Available Presets
============================================================

HV (High-Variance) Presets:
  • lightgbm
  • biglightgbm
  • xgboost
  • catboost
  • random_forest
  • decision_tree
  • linear_regression

LV (Low-Variance) Presets:
  • mlp
  • bigmlp
  • resnet
  • transformer

============================================================

Getting Preset Models Directly¶

You can also get preset models without using OGModel:

from og_learn.presets import get_hv_model, get_lv_model

# Get HV model
hv_model = get_hv_model('lightgbm')

# Get LV model (requires num_features)
lv_model = get_lv_model('mlp', num_features=10, epochs=100)

# Use directly
hv_model.fit(X_train, y_train)
predictions = hv_model.predict(X_test)