HAMLET: A framework for Human-centered AutoML via Structured Argumentation

Abstract

Machine Learning (ML) plays a crucial role in data analysis and data platforms (i.e., integrated sets of technologies that collectively meet end-to-end data needs). In the last decade, we have witnessed an exponential growth in both the complexity and the number of ML techniques; leveraging such techniques to solve real-case problems has become difficult for Data Scientists. Automated Machine Learning (AutoML) tools have been devised to alleviate this task, but easily became as complex as the ML techniques themselves — with Data Scientists losing again control over the process. In this paper, we design and extend HAMLET (Human-centered AutoMl via Logic and argumEnTation), a framework that helps Data Scientists to redeem their centrality. HAMLET enables Data Scientists to express ML constraints in a uniformed human- and machine-readable medium. User-defined constraints are interpreted to drive the exploration of ML pipelines (i.e., Data Pre-processing transformations shape the data so that the ML task can be performed at its best). AutoML retrieves the most performing pipeline instance, and finally, new constraints are learned and integrated through Logic and Argumentation. By doing so, HAMLET not only allows an easy exploitation of the knowledge acquired at each iteration, but also enables its continuous revision via the AutoML tool and the collaboration of both Data Scientists and domain experts.