Zoom Meeting: https://umass-amherst.zoom.us/j/96070156881?pwd=TnpVZjdtWlVGSWMrVklDV3BUY3Nkdz09
Abstract:The recent growth of data science expanded its reach to an ever-growing user base of non-experts, increasing the need for democratization and explainability in these systems. Democratization demands that a system can be used by people with different skills and backgrounds alike. Explainability helps the users understand and trust the system function, especially when unexpected behavior occurs. However, most existing data systems offer limited usability and support for explanations: these systems are usable only by experts with sound technical skills, and even expert users are hindered by the lack of transparency into the systems' inner workings and function. The aim of my thesis is to bridge the usability gap between non-expert users and complex data systems, and explain system behavior towards achieving trust while using these systems. To this end, my thesis has three goals: (1) enhancing usability of data systems for non-experts, (2) assisting users to achieve trust in data-driven machine learning, and (3) explaining system behavior to enable understanding of unexpected outcomes.
For enhancing usability, we focus on example-driven user intent discovery. We develop systems based on example-driven interactions in two different settings: querying relational databases and personalized document summarization. As a mechanism to achieve trust in machine learning, we develop a new data-profiling primitive that can characterize tuples for which a machine-learned model is likely to produce untrustworthy predictions. Finally, we develop explanation frameworks to explain causes of tuple non-conformance in the context of trusted machine learning, root cause of a concurrent application's intermittent failure, and cause of system malfunction due to certain data profiles.
Advisor: Alexandra Meliou
Committee members: Emery Berger, Peter J. Haas, Suman Nath