Super Learner: Stack Generalization Algorithm for AutoML
Mihir Gada, Zenil Haria, Arnav Mankad, Kaustubh Damania, Smita R. Sankhe · 2021
Model Selection and Ensembling are important and tedious tasks in the process of machine learning. To implement these processes, a lot of knowledge about machine learning as well as the application domain is needed. Often, many researchers perform trial and error methods to find the optimum model architecture for the application, and this process may take an indefinite amount of time. AutoML reduces the burden on the user and saves time, which can be used to invest in other important tasks like data gathering and analyzing the results produced. There are various state-of-the-art ensembling techniques like bagging, boosting, stacking, blending, etc. In this paper, we propose a variant of Super Learner which is a finely-tuned algorithm based on Stack Generalization. This algorithm is proposed for supervised learning on tabular-labeled datasets for prediction and classification tasks. We compare the performance of this ensemble with the traditional super learner algorithm as well as individual base models. These results show that the proposed variant of Super Learner outperforms the base learner models and the traditional Super Learner algorithm.