AMLA: Adaptive Meta-Learning Architecture for Automated Dataset Characterization, Predictive Algorithm Selection, and Feature Augmentation Advising

The adaptive process for choosing algorithms in applied machine learning identifies both labor and resource intensive bottlenecks. This paper describes a new architecture which automates algorithm suggestions for machine learning on structured tabular data. We present Adaptive Meta-Learning Architecture (AMLA), a domain-independent framework for each specific data set. AMLA contains a three-part Dataset Characterization Engine, Predictive Algorithm Selector, and Feature Augmentation Advisor. The Engine provides a multi-layered, 60-dimensional numerical fingerprint called “Dataset DNA,” as a result of the encoding of statistical, structural, information-theoretic, landmarking, and complexity features. The Selector is a trained meta-learner that explains the ranked algorithm suggestions based on the Dataset DNA, using SHAP (Shapley Additive Explanations) explanations

for pluses and minuses on algorithms. The Advisor identifies and suggests changes to the structure of the data set. Unlike other AutoML systems that function as black boxes, AMLA explains suggestions using an evolving Meta-Knowledge Base which includes data sets from OpenML community experiments and a local validation process. When tested using 50 benchmark classification sets, AMLA exceeded meta-learner Precision@1 of 72% which is a significant improvement over random and heuristic baselines. The completed system is hosted as a web application.

Message: Hi FE, I need help with “AMLA: Adaptive Meta-Learning Architecture for Automated Dataset Characterization, Predictive Algorithm Selection, and Feature Augmentation Advising” in “Computer Science & Engineering”