SUMO models

SUMo models compressed file is composed of two models: source separation and target separation. These models were trained using the OSTrain dataset (available as a Zenodo DOI at 10.5281/zenodo.8362616) using gradient boosting decision trees (xgboost library in Python). The source separation model was trained with the hyperparameters {colsample_bylevel = 0.7; colsample_bytree = 0.6; gamma = 0.1; learning_rate = 0.1; max_depth = 3; n_estimators = 800; reg_alpha = 1e-5; reg_lambda = 0.1}, and the target separation model was trained with the hyperparameters {colsample_bylevel = 0.4; colsample_bytree = 0.6; gamma = 0.1; learning_rate = 0.1; max_depth = 15; n_estimators = 800; reg_alpha = 1e-5; reg_lambda = 1}. These hyperparameters were obtained using Bayesian optimization and validated using the OSValidate dataset (available as a Zenodo DOI at 10.5281/zenodo.8360991).

Once decompressed, the source separation model will be at ./models/source_separation/source_separation_model_bayesian_optimization.joblib, and the target separation model will be at ./models/target_separation/target_separation_model_bayesian_optimization.joblib. The files at ./models/source_separation/log_parameters.txt and at ./models/target_separation/log_parameters.txt contain the log for all the hyperparameter combinations considered during the Bayesian optimization process.