DRIL: Descriptive Rules by Interactive Learning



Fang Cao, Eli T Brown

 External link (DOI) 

 View presentation:2020-10-30T14:10:00ZGMT-0600Change your timezone on the schedule page
2020-10-30T14:10:00Z

Exemplar figure — DRIL presents a technique using interactive machine learning to support a user in rapidly learning rules that describe data. By rapidly characterizing selected data in terms of what distinguishes them from the rest, DRIL facilitates rapid sensemaking that extends beyond the 'key features' of linear modeling, to provide an efficient overview of not only what features are important but what value ranges matter.

Fast forward

Direct link to video on YouTube: https://youtu.be/88f-JssKCSw

Keywords

human-in-the-loop, data characterization, rule learning

Abstract

Analyzing data is increasingly a part of jobs across industry, science and government, but data stakeholders are not necessarily experts in analytics. The human-in-the-loop (HIL) approach includes semantic interaction tools, which leverage machine learning behind the scenes to assist users with their tasks without engaging them directly with algorithms. One widely applicable model for how humans understand data is descriptive rules, which can characterize important attributes and simultaneously their crucial values or ranges. In this paper, we introduce an approach to help with data understanding via interactively and automatically generated rules. Our approach makes discerning the behavior of groups of interesting data efficient and simple by bridging the gap between machine learning methods for rule learning and the user experience of sensemaking through visual exploration. We have evaluated our approach with machine learning experiments to confirm an existing rule learning algorithm performs well in this interactive context even with a small amount of user input, and created a prototype system, DRIL (Descriptive Rules by Interactive Learning), to demonstrate its capability through a case study.