Foundations for Automated Data Science

Data science, despite its clear value, still has not received satisfactory formal treatment as a discipline. Many regard data science as a pragmatic black art, in large part due to the fact that data preparation, model deployment, and many practical model desiderata beyond simple predictive accuracy are generally not treated in courses or textbooks on statistics or machine learning in the sense of being guided by any rigorous underlying principles. This has resulted, for example, in much data science education focusing on the ability to use a collection of specific tools. It has also resulted in the widespread occurrence of subtle but significant conceptual errors being made in practice, even by PhDs in major institutions. In this talk I will present a mathematical model of data science that can clarify and guide the aforementioned important pragmatic aspects of data science rather than simply ascribing best practice to heuristics, general experience, or domain knowledge. I will discuss open practical issues in data science, including learnings from extensive user studies, show how such a theoretical foundation can address them, and finally show how these principles can translate to new practical data science tools in the form of the user experience, both graphical and programmatic in the form of libraries/languages.