AutoML Benchmark (AMLB)

AutoML frameworks compared on classification accuracy and ROC AUC across a shared set of OpenML tasks.

Results of the AutoML Benchmark (AMLB), which ran automated machine-learning frameworks on a shared set of classification tasks and published the evaluations on OpenML as study 226 (www.openml.org/s/226, 2019). Each subject is one framework; each run is one dataset, with the predictive accuracy and ROC AUC recorded there (0-1, higher is better).

Results

predictive_accuracy per subject over 6 measurements, ranked by average.

SubjectAvgMedianMinMaxP95P99Count
h2oautoml0.71490.71490.71490.71490.71490.71491
tunedrandomforest0.69220.69220.69220.69220.69220.69221
autosklearn0.68710.68710.68710.68710.68710.68711
randomforest0.67790.67790.67790.67790.67790.67791
tpot0.66770.66770.66770.66770.66770.66771
autoweka0.6630.6630.6630.6630.6630.6631

Metrics

  • predictive_accuracy — recorded on each measurement. Fraction of test instances classified correctly, from 0 to 1. Higher is better
  • area_under_roc_curve — recorded on each measurement. Area under the ROC curve, from 0 to 1. Higher is better; 0.5 corresponds to random guessing

Subjects (6)

  • autosklearn
  • autoweka
  • h2oautoml
  • randomforest
  • tpot
  • tunedrandomforest

Published by OpenML.