A Sequential Learning Framework with Applications to Learning from Crowds

While traditional machine learning usually deals with given static data, many online data are collected via a sequence of interactions with agents such as crowd labelers or customers. The motivating applications of the project include crowd labeling tasks (which is a powerful paradigm for utilizing human wisdom to collect data labels), sequential product recommendation, and online multi-product pricing. For all these applications, online learning and sequential decision-making are indispensable to each other. The objective of this project is to develop new sequential learning algorithms with rigorous theoretical guarantees. The developed framework will not only make fundamental technical contributions but also facilitate many important applications. For example, it will greatly improve the aggregated answers from crowd labelers with a significantly reduced cost. It can enhance the revenue of business while improving the customers’ satisfaction by providing accurate recommendations. In addition, this project also facilitates the development of new courses on machine learning for business school students, which helps bring the knowledge from data science to future business leaders, and provides training to K-12 students, with an emphasis on those from underrepresented groups.

This project strives to develop a unified learning and decision-making framework, which serves as an intellectual bridge connecting machine learning, stochastic optimization, and decision theory. In particular, there are three complementary research thrusts. The first thrust creates a suite of efficient algorithms that deal with complex task structures, such as ranking with transitivity structures or product recommendation with combinatorial structures, in a non-stationary environment. The algorithms will extend the bandit learning with finite independent arms into the setting with a complex correlation structure among potentially infinite number of arms. The second thrust seeks a cost-effective paradigm that either incorporates “optimal stopping” rule under a certain budget constraint or minimizes the sample complexity. The third thrust systematically evaluates the algorithms and theories on real problems coming from both crowdsourcing and other business-related applications. Moreover, since the computational efficiency and scalability is an important focus, the project will also advance the distributed statistical learning and stochastic optimization fields.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.

Project link: A Sequential Learning Framework with Applications to Learning from Crowds