Benchmark.

Ranking of the effects of the site (EoS)-removal effectiveness for different MAREoS.

Detection of complex effects including interactions.

Machine learning algorithm Observed extra accuracy
Lasso (glmnet) 0% (reference)
Lasso with first-order interactions (glmnet) +18%
Random forests (randomForest) +20%
Support vector machine (e1071) +16%
Gaussian processes (kernlab) +12%

Removal of simple EoS.

Name of MAREoS More information on the MAREoS Machine learning algorithm Average relative accuracy change in datasets with EoS (RACEoS) Average relative accuracy change in datasets with true effects (RACtrue) EoS-removal effectiveness
ComBat http://enigma.ini.usc.edu/
protocols/statistical-protocols/
Lasso (glmnet) 97% 0% 97%
Random forests (randomForest) 96% 0% 96%
Support vector machine (e1071) 92% 0% 92%
Gaussian processes (kernlab) 99% 0% 99%

Removal of complex EoS including interactions.

Name of MAREoS More information on the MAREoS Machine learning algorithm Average relative accuracy change in datasets with EoS (RACEoS) Average relative accuracy change in datasets with true effects (RACtrue) EoS-removal effectiveness
ComBat http://enigma.ini.usc.edu/
protocols/statistical-protocols/
Lasso with first-order interactions (glmnet) 0% 0% 0%
Random forests (randomForest) 84% 1% 84%
Support vector machine (e1071) 62% 1% 62%