Abstract by Brandon Schoenfeld
Machine Learning Pipeline Recommendation with Active Meta-learning
In the context of machine learning (ML), meta-learning is the use of previous ML experiments to inform current ML solutions. Past research has shown some success using ML to determine which classification algorithm will produce the best performance (e.g. most accurate) for a classification problem. By extracting characterizing information from datasets, called meta-features, a meta-model can learn a function, mapping meta-features to the best (known) classifier for that dataset and classification problem.
Almost always, before any classification algorithm can be applied to a dataset, the data must be preprocessed. Furthermore, some preprocessing steps can improve the performance of a chosen classifier. We extend the work of meta-learning to include the selection of a single preprocessing step to improve the performance of a chosen classifier, taking a step towards entire ML pipeline recommendation systems.