Model training implemented to be the same as Kaggle score=1.2

[kowaalczyk]

Implement training with default parameters same as in this kaggle kernel aiming to achieve similar feature importances. Do it for both XGB and CatBoost.

[kowaalczyk]

Status:

• LGBM performs a little better than XGB, but is significantly faster
• We have run the GridSearch to find the optimal params on a simple aggregations dataset
• Need to move LGBM implementation to a script and fix XGBoost training script in scripts.py

[kowaalczyk]

Best LGB params:

boosting_type=gbdt, colsample_bytree=0.5, learning_rate=0.03, max_depth=5, metric=multi_logloss, min_child_weight=10, min_split_gain=0.01, n_estimators=1000, num_class=14, objective=multiclass, reg_alpha=0.001, reg_lambda=0.1, silent=-1, subsample=0.9, verbose=-1