Title: Empowering non-experts communicate preferences to machine learning models in Visual Analytics

Subhajit Das

Ph.D. Student

School of Interactive Computing

Georgia Institute of Technology 

Date: Friday, September 27th, 2019

Time: 04:00 pm – 06:00 pm (EST)

Location: TSRB 223 (2nd floor)

Committee:

Dr. Alex Endert  (Advisor, School of Interactive Computing, Georgia Institute of Technology)

Dr. John Stasko (School of Interactive Computing, Georgia Institute of Technology)

Dr. Polo Chau  (School of Computational Science and Engineering, Georgia Institute of Technology)

Dr. Thomas Ploetz,  (School of Interactive Computing, Georgia Institute of Technology)

Dr. Remco Chang (Department of Computer Science, Tufts University)

Abstract:

Recent visual analytic (VA) systems rely on machine learning (ML) to allow users perform a variety of data analytic tasks, e.g., biologists clustering genome samples, medical practitioners predicting the diagnosis for a new patient etc. Such VA systems support interactive construction of machine learning models to people who are not formally trained in ML or data science (i.e., non-experts in ML). However, designing VA systems for non-experts poses various challenges. For example, conventionally ML models are constructed by ML practitioners or data science experts who know model parameters/hyperparameters, model metrics, and the underlying theory that drive several algorithms in machine learning.  On the other hand, non-experts are often overwhelmed by such complexity, thus inhibiting their access to various ML-based processes that can improvise their current data analysis goals. Existing VA systems that support interactive model construction for non-experts still require considerable data science expertise or an ML practitioner to communicate their preferences to these model(s). To solve this gap, we need VA systems which provide novel interactive techniques for non-experts to communicate their choices/preferences to adjust the underlying model(s); without the need to understand model parameters, hyperparameters, complex metrics, or other related data science expertise.  Through my research, I seek to investigate how non-experts in ML can​ communicate their preferences to interactively construct models. Specifically, I explore and analyze several VA techniques such as multi-model steering, model selection, interactive objective functions, etc. to facilitate specification of user goals and objectives to the underlying model(s) in VA systems.