Copyright 2019 Reflux Design
Bio
I am an Assistant Research Professor at the Center for Data Science for Enterprise and Society (CDSES) in Cornell University. I am interested in data-driven decision-making for business problems, especially in marketing, such as advertising, recommendation, promotions and pricing. Specifically, my current research focuses on online learning and optimization problems in digital marketing. I have also worked on approximation algorithms for discrete optimization in the past.
A large bulk of of my work are motivated by collaboration with industry. For example, we collaborated with Glance, India's largest mobile lock-screen content platform, to improve their recommender system for short-lived contents, and showed the effectiveness of our policy in a large-scale field experiment.
Prior to joining Cornell, I obtained my PhD degree in the ACO program (Algorithms, Combinatorics and Optimization) at the Tepper School of Business, Carnegie Mellon University. I was fortunate to be able to work with R. Ravi and Andrew Li. I received my BS degree in Mathematics from Tsinghua University in 2014, and MS degree in Applied Mathematics and Statistics from the State University of New York at Stony Brook in 2017, where I was fortunate to work on computational geometry with Joseph Mitchell and Jie Gao.
I received the INFORMS Pierskalla Best Paper Award in Health Applications in 2021 and the INFORMS George B. Dantzig Dissertation Award Dissertation Award in Operations Research and Management Science in 2022. Here is my CV.
Selected Work
The full list of my papers can be found in a separate section.
Short-Lived High-Volume Bandits: Algorithms and Field Experiment
Su Jia, Nishant Oli, Andrew Li, R. Ravi, Paul Duff, Ian Anderson.
Platforms have been leveraging the scale of data to build recommenders for newly-generated contents. Drawing ideas from bandit theory, we propose policies for recommending a high-volume of short-lived contents, which we also show to be nearly best possible theoretically. More importantly, we collaborated with Glance, India's largest lockscreen content company and implmented a large-scale field experiment on their real system, where our policy significantly outperformed their current DNN-based recommender.
Active Sequential Hypothesis Testing
In clinical trials, tests are performed sequentially to identify a patient's unknown disease. In marketing, AB tests are performed sequentially to determine the best design of a website. We study the design of low-cost test procedures given a set of tests.
Unknown Outcomes: Optimal Decision Tree and Submodular Ranking with Noisy Outcomes. Su Jia, Fatemeh Navidi, Viswanath Nagarajan and R.Ravi. (NeurIPS'19)
Noisy Outcomes: Towards A Liquid Biopsy: Approximation Algorithms for Active Sequential Hypothesis Testing. Kyra Gan, Su Jia, Andrew Li, Sridhar Tayur. (NeurIPS'21) Winner, INFORMS Pierskalla Best Paper Award 2021
Dyanmic Pricing Under Monotonicity Constraint
While well-understood theoretically, bandit-based pricing policies are rarely deployed in the real world, largely for overlooking practical constraints. For example, oscillating prices may be unfavorable in many applications. We consider policies with non-increasing prices, and show optimal regret bounds that "separate" from the unconstrained setting.
Non-parametric model: Conservative Price Experiments: Markdown Pricing Under Unknown Demand. Su Jia, Andrew Li and R.Ravi. (Major Revision, Management Science)
Parametric model: Markdown Pricing Under Unknown Parametric Demand Models. Su Jia, Andrew Li, R. Ravi (NeurIPS'22)
Effective Online Order Acceptance Policies for Omni-Channel Fulfillment.
Su Jia, Jeremy Karp, R. Ravi, Sridhar Tayur. (Manufacturing and Service Operations Management)
Retailers nowadays may use in-store inventory to fulfill the demands from different channels. As opposed to the offline channel, the online channel is more flexible and its orders are fulfilled only periodically. The key problem then is how to jointly manage the orders from these two channels, with consideration of both profit and fulfillment level. We propose a gradient-based computational framework for finding the optimal threshold policy, and demonstrate its effectiveness using Onera's data.