Analysis and Design of Micro-Task Crowdsourcing Platforms
Speaker: Djellel Difallah, NYU Center for Data Science
Location: 60 Fifth Avenue 527
Date: June 3, 2019, 11 a.m.
Host: Lakshmi Subramanian
Existing micro-task crowdsourcing platforms, such as Amazon Mechanical Turk, are often treated as a black box for on-demand human intelligence. In reality, such platforms are marketplaces with complex dynamics that are important to study. On the one hand, workers have different expertise levels, motivating incentives and demographics. On the other hand, multiple requesters with varying workloads and performance requirements compete for the attention of the available workforce. In this talk, I will present the results of longitudinal analyses of data collected from Amazon Mechanical Turk using web scraping and surveys. The goal of these analyses is to gain insight into the underlying dynamics of the crowd and to guide the design of more effective and efficient crowdsourcing platforms. Next, I will present the architecture of a push-based system that supports fairness among production and best effort tasks. Finally, I will discuss a probabilistic model for estimating the crowd size; This estimate is particularly relevant for designing future crowdsourcing systems that can offer deadline guarantees.
Djellel Difallah is a Moore-Sloan Fellow at the NYU Center for Data Science. His current research interests revolve around studying the dynamics of the crowd in micro-task crowdsourcing systems and collaborative knowledge bases. He received a Ph.D. from the University of Fribourg (Switzerland) in 2015. Prior to that, he received a Dipl.-Ing. in 2004 from the University of Science and Technologies in Algiers, and an M.S. in Computer Sciences in 2011 from the University of Lousiana at Lafayette where he was a Fulbright scholar. Djellel occupied various engineering roles in the telecommunication and the E&P industries, and during his graduate studies, he interned at a Google-sponsored program and at Microsoft Research.
Refreshments will be offered starting 15 minutes prior to the scheduled start of the talk.