Abstract
We introduce an automatic machine learning (AutoML) modeling architecturecalled Autostacker, which combines an innovative hierarchical stackingarchitecture and an Evolutionary Algorithm (EA) to perform efficient parametersearch. Neither prior domain knowledge about the data nor feature preprocessingis needed. Using EA, Autostacker quickly evolves candidate pipelines with highpredictive accuracy. These pipelines can be used as is or as a starting pointfor human experts to build on. Autostacker finds innovative combinations andstructures of machine learning models, rather than selecting a single model andoptimizing its hyperparameters. Compared with other AutoML systems on fifteendatasets, Autostacker achieves state-of-art or competitive performance both interms of test accuracy and time cost.