Abstract by Brandon Schoenfeld
A Dynamic Neural Architecture (DNA) for Transferable Metalearning over Machine Learning Pipelines
State-of-the art automatic machine learning (AutoML) systems produce predictive models (pipelines) which include data cleaning, preprocessing, feature selection, modeling, ensembling, etc. With a combinatorial number of possible solutions, AutoML systems typically restrict the solution space, utilize past experience (metalearning), and/or employ other heuristics to find quality solutions in a reasonable time frame. We propose a metalearning approach which could enable an AutoML system to estimate pipeline performance in a much less restricted space. This approach utilizes a novel neural network architecture for directed acyclic graphs (DAGs) and apply it to quickly estimate pipeline performance. Our meta-model dynamically modifies its neural architecture to mimic the structure of any pipeline (a DAG).