Seminars / Conferences

We present a novel podcast recommender system deployed at industrial scale. This system successfully optimizes personal listening journeys that unfold over months for hundreds of millions of listeners. In deviating from the pervasive industry practice of optimizing machine learning algorithms for short-term proxy metrics, the system substantially improves long-term performance in A/B tests. The paper offers insights into how our methods cope with attribution, coordination, and measurement challenges that usually hinder such long-term optimization. To contextualize these practical insights within a broader academic framework, we turn to reinforcement learning (RL). Using the language of RL, we formulate a comprehensive model of users’ recurring relationships with a recommender system. Then, within this model, we identify our approach as a policy improvement update to a component of the existing recommender system, enhanced by tailored modeling of value functions and user-state representations. Illustrative offline experiments suggest this specialized modeling reduces data requirements by as much as a factor of 120,000 compared to black-box approaches.

Time permitting, I will also discuss a novel exploration problem that arises from this work, in which one is tasked with quickly evaluating new pieces of content even though the true (long-term) reward metric takes months to realize.

*This talk is based on joint work with Lucas Maystre and Yu Zhao, from Spotify.

While labor market inequality research has extensively explored racial and gender disparities in hiring, wages, benefits, and promotions, the inequality in the temporal quality of work schedules has largely been overlooked, despite its significant economic and social consequences for workers. Using large and granular shift data from an American casual dining chain, we examine racial and gender disparities in scheduling quality across two key dimensions: (1) sufficiency, which assesses the number of hours workers receive, and (2) predictability, which examines the frequency of last-minute shift changes. We address two main questions: (1) Are there gender and racial disparities in scheduling quality, and if so, what are they? (2) What mechanisms drive these disparities? Our findings reveal that women are initially scheduled for up to 11.7% fewer hours than men but experience more frequent additions after accounting for employment status, experience, and skill levels. We also show that alternative explanations, such as women’s lower availability, preference for flexibility, and reliability, either do not hold true or only explain a small portion of the observed disparities. Thus, we provide compelling evidence for the presence of managerial bias in scheduling decisions. By leveraging detailed field data, analyzing multiple dimensions of schedule quality, and systematically examining alternative mechanisms, we deepen the understanding of the drivers behind the gender disparities in schedule quality and offer important managerial implications for addressing these issues.

Recidivism, where formerly incarcerated individuals reoffend, is a significant challenge in the U.S. criminal justice system with an over 50% reoffending rate. Incarceration diversion programs aim to break this revolving door by focusing on rehabilitation rather than incarceration. Machine learning (ML) algorithms assist in admission decisions to these programs by assigning recidivism risk scores to potential participants, helping to balance the program benefits against the risk of in-program reoffense. However, ML risk estimation can be imperfect and biased, leading to potentially suboptimal and unfair admission decisions. We formulate an admission control problem using a queueing model with heterogeneous classes and covariate-based risk estimation. Our model accounts for in-service reneging, where participants may reoffend before completing the program. We derive a priority score policy from a static control problem, analyze the performance of this priority policy under imperfect estimation, and propose conditions for its optimality. Additionally, we leverage Markov Decision Processes to develop dynamic decision support. If time permits, I will also discuss some ongoing work that leverages generative AI models to assist in identifying individuals at risk and our engagement with community partners to deploy the decision support tools. This talk is based on joint work with Xiaoquan Gao, Bingxuan Li from Purdue University, and Zhiqiang Zhang, Amy Ward from the University of Chicago.

Generative artificial intelligence (AI) is poised to revolutionize how humans work, and has already demonstrated promise in significantly improving human productivity. However, a key remaining question is how generative AI affects learning, namely, how humans acquire new skills as they perform tasks. This kind of skill learning is critical to long-term productivity gains, especially in domains where generative AI is fallible and human experts must check its outputs. We study the impact of generative AI, specifically OpenAI’s GPT-4, on human learning in the context of math classes at a high school. In a field experiment involving nearly a thousand students, we have deployed and evaluated two GPT based tutors, one that mimics a standard ChatGPT interface (called GPT Base) and one with prompts designed to safeguard learning (called GPT Tutor). These tutors comprise about 15% of the curriculum in each of three grades. Consistent with prior work, our results show that access to GPT-4 significantly improves performance (48% improvement for GPT Base and 127% for GPT Tutor). However, we additionally find that when access is subsequently taken away, students actually perform worse than those who never had access (17% reduction for GPT Base). That is, access to GPT-4 can harm educational outcomes. These negative learning effects are largely mitigated by the safeguards included in GPT Tutor. Our results suggest that students attempt to use GPT-4 as a “crutch” during practice problem sessions, and when successful, perform worse on their own. Thus, to maintain long-term productivity, we must be cautious when deploying generative AI to ensure humans continue to learn critical skills.

In the first part of this talk, I will talk about the idea of meta-learning by going over a few representative earlier studies, including neural processes, learned optimizers, hyperparameter tuning and in-context learning, in order to motivate audience to think of neural net learning from a broader perspective. I will then talk about how this perspective allows us to think of neural net learning as a way to learn an algorithm, enabling us to arrive at algorithms we could not have ourselves. In the second part of my talk, I will describe two ongoing studies on causal discovery and causal inference.

Ride-hailing platforms face two important challenges: (i) there are significant spatial demand imbalances that require some repositioning (empty routing) of drivers; (ii) the control of driver supply is partially decentralized in that drivers strategically decide whether to join the network, and if so, whether and where to reposition when not serving riders.

We study the following question for such ride-hailing networks: Under decentralized repositioning, how effective are driver wage policies in achieving the optimal centralized performance benchmark?
We consider a stationary fluid model of a ride-hailing network in a game-theoretic framework with riders, drivers, and the platform. We show how the effectiveness of driver wage policies under decentralized repositioning depends on the interplay of the network’s spatial (travel time) configuration, driver wage flexibility, and the congestion-sensitivity of travel times: (1) We identify conditions on the travel times for the existence of a driver repositioning equilibrium. (2) For networks with constant travel times, we show that the centrally optimal repositioning flows can be implemented under decentralized repositioning, provided the platform has sufficient wage flexibility, whereas more limited wage flexibility leads to inefficiencies in terms of driver idling. (3) For networks with congestion-sensitive travel times, the centrally optimal repositioning flows can generally not be implemented under decentralized repositioning, even with full wage flexibility, so that decentralized repositioning leads to higher congestion, capacity levels, and driver wage rates compared to centralized repositioning.

(Joint work with Andre Cire and Uta Mohring.)