Explore-By-Example

Abstract

Traditional DBMSs are suited for applications in which the structure, meaning and contents of the database, as well as the questions (queries) to be asked, are all well-understood. However, this is no longer true when the volume and diversity of data grow at an unprecedented rate, while the user ability to comprehend data remains (as limited) as before. To address the increasing disparity in the data - same humans problem, our project explores a new approach of system-aided exploration of a big data space and automatic learning of the user interest in order to retrieve all objects that match the user interest -- we call this new service data which complements the traditional querying interface of a database system.In this talk, I introduce a new framework for interactive data exploration, called Explore-by-Example, which iteratively seeks user relevance feedback on database samples and uses such feedback to finally predict a query that retrieves all objects of interest to the user. The goal is to make such exploration converge fast to the true user interest model, while minimizing the user labeling effort and providing interactive performance in each iteration. I discuss a range of techniques and optimizations to do so for linear patterns and complex non-linear patterns. Our user study indicates that our approach can significantly reduce the user effort and the total exploration time, compared with the common practice of manual exploration. I finally conclude the talk by pointing out a host of new challenges, ranging from application of active learning theory, to database optimizations, to visualization.