params.txt

# Collected examples of LGB and XGB params across Kaggle, etc that worked so as to get a "methodical" feel for tuning.

# XGB Example Params 
# =============================

params = {'eta': 0.09, 
'max_depth': 4, 
'objective': 'binary:logistic', 
'subsample': 0.8, 
'colsample_bytree': 0.8, 
'min_child_weight': 0.77, 
'scale_pos_weight': 1.6, 
'gamma': 10, 
'reg_alpha': 8, 
'reg_lambda': 1.3, 
'eval_metric': 'auc', 
'seed': 99, 
'silent': True}

oliver={'alpha': [0, 0.01, 100],
 'booster': ['gbtree'],
 'colsample_bytree': [0.8, 0.6],
 'early_stopping': [50],
 'eta': [0.1],
 'eval_metric': ['mlogloss'],
 'feval': [xgb_gini],
 'gamma': [1, 100],
 'lambda': [1, 100],
 'max_depth': [4, 2, 6, 8, 10, 12],
 'maximize_feval': [True],
 'min_child_weight': [8],
 'num_class': [2],
 'num_estimators': [1000000000],
 'objective': ['multi:softprob'],
 'scale_pos_weight': [26],
 'silent': [0],
 'subsample': [0.8, 0.6]}

{'gamma': 1, 'max_depth': 6, 'subsample': 0.8, 'feval': <function xgb_gini at 0x7f981747e598>, 'maximize_feval': True, 'min_child_weight': 8, 'booster': 'gbtree', 'num_class': 2, 'colsample_bytree': 0.6, 'alpha': 0, 'objective': 'multi:softprob', 'num_estimators': 1000000000, 'eta': 0.1, 'early_stopping': 50, 'lambda': 100, 'eval_metric': 'mlogloss', 'scale_pos_weight': 26, 'silent': 0}

{'max_depth': 3.0, 'gamma': 9.0, 'min_child_weight': 12.0, 'max_delta_step': 2.0, 'subsample': 0.80000000000000004, 'colsample_bytree': 0.80000000000000004}

# LGB Example Params 
# =============================

params = {'learning_rate': 0.02, 
'max_depth': 4, 
'boosting': 'gbdt', 
'objective': 'binary', 
'max_bin': 10, 
'subsample': 0.8, 
'subsample_freq': 10, 
'colsample_bytree': 0.8, 
'min_child_samples': 500, 
'metric': 'auc', 
'is_training_metric': False, 
'seed': 99}

lgbparam = {'boosting_type':['gbdt'],
 'num_estimators': [5000],
 'is_unbalance': [False],  
 'num_leaves': [8, 4, 12, 16, 32, 256, 512], # number of leaves in one tree
 'early_stopping': [100], 
 'learning_rate': [0.05], 
 'min_data_in_leaf': [25, 500], # min_child_samples
 'use_missing': [True], # -1
 'num_threads': [3], 
 'objective': ['binary'],
 'feature_fraction': [0.8, 0.3], # colsample_bytree
 'predict_raw_score': [False], 
 'bagging_freq': [5, 10, 2], # subsample_freq
 'bagging_fraction': [0.8, 0.7], # subsample
 'lambda_l2': [0.0],
 'max_bin':[255, 10, 20], # def 255
 'feval':[None],
 'maximize_feval':[True],
 'num_class':[1],
 'max_depth':[-1, 4, 6], # def -1
 'min_hessian':[1e-3, 0.05], # min_child_weight def 1e-3
 'eval_metric':[['binary_logloss', 'auc']]}