Multi-stage heterogeneous ensemble meta-learning with hands-off user-interface and stand-alone prediction using principal components regression: The R package EnsemblePCReg

Despite the fact that ensemble meta-learning of a heterogeneous collection of base learners is an effective means to reduce the generalization error in predictive models, several factors have impeded a broad adoption of such techniques among practitioners. These factors include an intractable number of choices of base learners and their tuning parameters, complex methodology required for integration of base learners, the ensuing complexity of software needed to support stand-alone prediction, and significant CPU and memory consumption of heterogeneous ensemble meta-learning techniques. The R package EnsemblePCReg overcomes the above barriers by combining several features. Sensible base-learner parameter grids provide a hands-off API for non-experts while allowing expert users to exert control by overriding default settings. Sophisticated ensemble generation and integration methods, combining stacked generalization and principal components regression, offer favorable generalization performance. Finally, computational optimizations such as advanded thread scheduling for improved parallelization scaling and file methods for relieving memory consumption during training and prediction, significantly increase the range of data sizes that can be handled on personal computers. In combining these features, EnsemblePCReg significantly lowers the barrier for practitioners to apply heterogeneous ensemble meta-learning techniques to their everyday regression problems.