Can't understand what's going on with LightGBM (Windows platform). Previously I had this algorithm really powerful, but now his performance is so bad.
For comparison (default parameters in each algorithm) LightGBM performs according to simple DIFF-metric = (actual - prediction):
I was trying to find some better parameters with HyperOpt, but also without success
LGBM_SPACE = {
'type': 'LGBM',
'task': hp.choice('lgbm_task', ['train', 'prediction']),
'boosting_type': hp.choice('lgbm_boosting_type', ['gbdt', 'dart']),
'objective': hp.choice('lgbm_objective', ['regression']),
'n_estimators': hp.choice('lgbm_n_estimators', range(10, 201, 5)),
'learning_rate': hp.uniform('lgbm_learning_rate', 0.05, 1.0),
'num_leaves': hp.choice('lgbm_num_leaves', range(2, 7, 1)),
'tree_learner': hp.choice('lgbm_tree_learner', ['serial', 'feature', 'data']),
'metric': hp.choice('lgbm_metric', ['l1', 'l2', 'huber', 'fair']),
'huber_delta': hp.uniform('lgbm_huber_delta', 0.0, 1.0),
'fair_c': hp.uniform('lgbm_fair_c', 0.0, 1.0),
'max_depth': hp.choice('lgbm_max_depth', range(3, 11)),
'min_data_in_leaf': hp.choice('lgbm_min_data_in_leaf', range(0, 6, 1)),
'min_sum_hessian_in_leaf': hp.loguniform('lgbm_min_sum_hessian_in_leaf', -16, 5),
'feature_fraction': hp.uniform('lgbm_feature_fractionf', 0.0, 1.0),
'feature_fraction_seed': hp.choice('lgbm_feature_fraction_seed', [12345]),
'bagging_fraction': hp.uniform('lgbm_bagging_fraction', 0.0, 1.0),
'bagging_freq': hp.choice('lgbm_bagging_freq', range(0, 16, 1)),
'bagging_seed': hp.choice('lgbm_bagging_seed', [12345]),
'min_gain_to_split': hp.uniform('lgbm_min_gain_to_split', 0.0, 1.0),
'drop_rate': hp.uniform('lgbm_drop_rate', 0.0, 1.0),
'skip_drop': hp.uniform('lgbm_skip_drop', 0.0, 1.0),
'max_drop': hp.choice('lgbm_max_drop', [-1] + range(2, 51, 1)),
'drop_seed': hp.choice('lgbm_uniform_drop', [12345]),
'verbose': hp.choice('lgbm_verbose', [-1]),
'num_threads': hp.choice('lgbm_threads', [2]),
}
The best result was just 450422301, that is super bad in comparing with above.
Example of using as all scikit-learn API:
model = LGBMRegressor()
model.fit(X, Y)
model.predict(XT)
Please try to use the latest code from master branch. There was an occurrence of inconsistent parameters in Scikit-learn API which was fixed: #1033.
Or you could add to your alg_conf "min_child_weight": 1e-3, "min_child_samples": 20.