Automated Machine Learning
Compete Mode
Achieve the best predictive accuracy with AutoML.
State-of-the-art performance .
Achieve the best predictive accuracy with AutoML.
State-of-the-art performance .
The mljar AutoML framework can work in
the Compete mode.
The Compete mode is perfect to get the best performing ML pipeline.
The Compete mode provides the high accuracy ML pipeline. It is a perfect mode for solutions where small percent of improvemnt has big impact. The example use cases for Compete mode are trading or data science competitions.
Use advanced feature engineering techniques to improve the performance. Enhance your data with Golden Features, K-Means Features. Use Feature Selection to train ML models only on relevant features.
The Compete mode use many different ML algorithms.
It ensembles and stacks them to get the best score.
Properly implemented to avoid overfitting.
The Compete mode uses the following Machine Learning algorithms:
Decision Tree, Random Forest, Extra Trees, Xgboost, LightGBM, CatBoost,
Neural Network, Nearest
Neighbors, Ensemble, Stacked Ensemble.
Just set mode="Compete"
during AutoML initialization and you are set!
"""AutoML Compete code """
from supervised import AutoML
# Initialize AutoML in Compete Mode
automl = AutoML(mode="Compete")
automl.fit(X, y)
The validation strategy is set automatically based on time budget and machine speed.
The validation can be: 75%/25%
train/test split, 5-fold Cross-Validation or 10-fold Cross
Validation.
The AutoML can detect the type of Machine Learning task to be solved based on target feature values. The supported ML tasks are: binary, multi-class classification, and regression.
Based on selected ML task type the proper metric for optimization will be choosen.
The classification tasks will optimize LogLoss
, while the regression tasks will minimize RMSE
.
In the Compete mode the AutoML will use feature engineering techniques like: Golden Features, K-Means Features. It improves the final model performance.
Tune ML models by checking different set of hyperparameters. The AutoML tune ML algorithms in three steps. Firstly, it checks the performance with a default set of HP. Then do a random-search over a defined set of HP. In the last step, it uses hill-climbing to fine-tune HP.
In this technique, the ML model is trained with sample weights boosted on previous best models errors. It is very similar to AdaBoost algorithm, except that there are only two iterations (one model with sample weights all equal 1, and second model with boosted sample weights on previous model errors).
The Compete uses Ensemble Averaging and Ensemble Stacking methods.
They are only available for Cross-Validation.
You can further customize Compete settings. You can easily switch on/off selected procedures by setting proper variable during AutoML initialization.