Novel uses of provenance in data science applications

:Abstract

Provenance – the “why” and “where” of data – has been extensively studiedand has been used to understand the results of database queries, debug networks and analyze data science workflows In this talk I will discuss two novel uses of data provenance:  creating fine-grained citations for data extracted from a database; and incrementally updating machine learning models after deletions have been made to the training set.  I will also show how the latter can be used to efficiently clean label uncertainties in machine learning training sets