AutoMLBench

This project is maintained by DataSystemsGroupUT

AutoMLBench: A Comprehensive Experimental Evaluation of Automated Machine Learning Frameworks

In this study, we benchmark the most commonly used tools, i.e.AutoSKLearn,TPOT, ATM, Recipe, SmartML, AutoWeka. We analyze the underlying techniques, and experimentally study the promising areas for each tool. For instance, the effect ofmeta-learning, ensembling, time budget, search space size and robustness of the optimization process have been empiricallystudied. The statistical significance of the accuracy differenceusing these techniques has been evaluated using Wilcoxontest. The results from 100 datasets show that the ensembling mechanism generally enhance the performance accuracy while the meta-learning mechanism is effective with very short time budgets only.