Doing more with less

Automated machine learning (AutoML) promises to democratize by automatically generating pipelines with little no user intervention. Typically, a search procedure is used repeatedly generate and validate candidate pipelines, maximizing predictive performance metric, subject limited execution time budget. While this approach candidates works well for small tabular datasets, the same does not directly scale larger datasets 100,000s of observations, often producing fewer yielding lower performance, given We carry out an extensive empirical evaluation impact that downsampling - reducing number rows in input dataset has on produced genetic-programming-based AutoML classification tasks.