While experimenting with MLJar, I figured out in 'Compete' mode it uses the below 3 steps towards the end of the training:
stack
ensemble
stack ensemble
As stacking is a type of ensemble learning, how are the last 3 steps different? I tried reading the documentation but no specifics are mentioned: https://supervised.mljar.com/features/automl/#ensemble
The description of each ensemble type in MLJAR AutoML package is in Algorithms section in the docs.
The docs for Ensemble is here. It is simple average of previous models. Models are selected until they improve the ensemble performance.
The docs for Stacked Algorithm is here. It is a model trained on original data plus stacked predictions of previous models.
The Stacked Ensemble is the Ensemble build from models trained on original data and models on stacked data (original+stacked predictions).
@mehul-gupta please let me know if it is clear now.