Workshop on Decision Intelligence and Analytics for Online Marketplaces:
Jobs, Ridesharing, Retail, and Beyond
August 15, 2022 Washington, DC
Keynote speakers are listed in alphabetical order of the last names.
Susan Athey
The Economics of Technology Professor, Stanford
Professor Athey is the Economics of Technology Professor at Stanford Graduate School of Business. She received her bachelor’s degree from Duke University and her PhD from Stanford, and she holds honorary doctorates from Duke University and London Business School. She previously taught at the economics departments at MIT, Stanford, and Harvard. She is an elected member of the National Academy of Science and is the recipient of the John Bates Clark Medal, awarded by the American Economics Association to the economist under 40 who has made the greatest contributions to thought and knowledge. Her current research focuses on the economics of digitization, marketplace design, and the intersection of econometrics and machine learning. She has worked on several application areas, including timber auctions, internet search, online advertising, the news media, and the application of digital technology to social impact applications. As one of the first “tech economists,” she served as consulting chief economist for Microsoft Corporation for six years, and has served on the boards of multiple private and public technology firms. She also served as a long-term advisor to the British Columbia Ministry of Forests, helping architect and implement their auction-based pricing system. She was a founding associate director of the Stanford Institute for Human-Centered Artificial Intelligence, and she is the founding director of the Golub Capital Social Impact Lab at Stanford GSB. She is currently on partial leave from Stanford to serve as the chief economist at the antitrust division of the U.S. Department of Justice.
Title:
Improving Fairness and Efficiency Using Estimates of User Preferences for Profile Characteristics in Online Marketplaces
Abstract:
In this paper, we consider the problem faced by an online marketplace seeking to balance fairness and efficiency in an environment where platform participants may have preferences characteristics of potential partners that may be revealed in photographs. We consider the setting of a microlending marketplace, Kiva, where lenders select among borrowers based on the borrower profiles. We begin by using observational data to construct estimates of the effect of features detected in images on funding rates and defaults. We find that lenders prefer images that depict female borrowers, smiling borrowers, and that are not body shots; but these characteristics are not strongly correlated with repayment rates. Although our estimation approach adjusts flexibly for a range of features detected using state-of-the-art computer vision techniques, there is still potential for bias due to information that lenders might be able to discern from photographs that we do not capture.
To address the potential for confounding, we conduct a survey experiment with recruited subjects to isolate the impact of image characteristics on lender choices. Experimental subjects choose between fabricated profiles generated using Generative Adversarial Networks that differ in the chosen features. The results of the experiment are consistent with the observational estimates. Finally, we observe that low-performing borrowers and females are more likely to smile and less likely to use body shots, and these choices exacerbate the gender gap on the platform. We simulate outcomes under various counterfactual policies aimed at improving fairness and efficiency by encouraging different choices about profiles. We show that policies based on image profile recommendations can boost outcomes of the least popular campaigns and reduce overall inequity without sacrificing the number of transactions.
Abraham Bagherjeiran
Senior Director, Applied Research at eBay
Abraham is a Sr. Director of Applied Researcher at eBay where he leads search ranking and monetization. Abraham has been in the ecommerce and computational advertising space for over 10 years. He has authored dozens of papers in the field and holds several related patents. He is also a part-time lecturer of Computer Science at the Northeastern University.
Title: Empowering people and creating economic opportunity through machine learning at eBay
Abstract: In over 190 markets, eBay helps millions buyers find unique items from small and large sellers across the world. Empowering buyers to find that perfect item among billions of possibilities anywhere in the world is a difficult challenge. In this talk we will discuss some of eBay's uniquely challenging problems and some of the innovative machine learning solutions we have built to address them.
Ido Bright
Staff Research Scientist, Lyft
Ido Bright is a Staff Data Scientist at Lyft, working on various aspects of ride-share from dispatch, driver positioning, passenger-matching and forecasting. His interests are in combining techniques from machine learning optimization and experimentation to solve real world problems. Prior to that he conducted his post-doc research at the University of Washington focusing on optimization and applying machine learning methods to partial differential equations. He received his PhD from the Weizmann Institute of Science focusing his thesis on averaging and infinite horizon optimization.
Title: Reducing Marketplace Interference Bias Via Shadow Prices
Abstract: Marketplace companies rely heavily on experimentation when making changes to the design or operation of their platforms. The workhorse of experimentation is the randomized controlled trial (RCT), or A/B test, in which users are randomly assigned to treatment or control groups. However, marketplace interference causes the Stable Unit Treatment Value Assumption (SUTVA) to be violated, leading to bias in the standard RCT metric. In this talk, we propose a technique for platforms to run standard RCTs and still obtain meaningful estimates despite the presence of marketplace interference. We specifically consider a matching setting, in which the platform explicitly matches supply with demand via a matching algorithm. Our proposed technique is quite simple: instead of comparing the total value accrued by the treatment and control groups, we instead compare each group’s average shadow price in the matching linear program. At the heart of our result is the idea that it is relatively easy to model interference in matching-driven marketplaces since, in such markets, the platform intermediates the spillover.
We prove that, in the fluid limit, our proposed technique corresponds to the correct first-order approximation (in a Taylor series sense) of the value function of interest. We then use this result to prove that, under reasonable assumptions, our estimator is less biased than the RCT estimator. We also present an extension of this result to supply-chains.
Daniel Hewlett
Senior Staff Machine Learning Engineer, LinkedIn
Bio: Daniel Hewlett is a Senior Staff AI Engineer at LinkedIn. He is a technical lead supporting LinkedIn Talent Solutions, leading the development of foundational AI technology to support LinkedIn’s applications for job seekers, recruiters, and learners. His interests include Multi-Task and Transfer Learning, Natural Language Understanding, and Learning to Rank. Previously, Daniel worked in the NLP group at Google Research, and at YouTube, after receiving his Ph.D. in Computer Science from the University of Arizona in 2011.
Title: Optimizing Hiring Outcomes on LinkedIn
Abstract: LinkedIn Talent Solutions provides a variety of applications to job seekers and hirers, including Job Search and LinkedIn Recruiter. In this talk, we describe how we have designed our AI systems to support the full range of job and member matching problems across search and recommendations, and both sides of the hiring marketplace. We discuss how we measure hiring outcomes and enable our models to effectively optimize hiring outcomes. We use transfer learning to provide ML models across all our applications with a shared understanding of member and job content such as profiles and job postings, as well as job-seeking activities.
Haixun Wang
VP Engineering, Distinguished Scientist, Instacart
Haixun Wang is an IEEE fellow, editor in chief of IEEE Data Engineering Bulletin, and a VP of Engineering and Distinguished Scientist at Instacart. Before Instacart, he was a VP of Engineering and Distinguished Scientist at WeWork, a Director of Natural Language Processing at Amazon, and he led the NLP team working on Query and Document Understanding at Facebook. From 2013 to 2015, he worked on natural language processing with Google Research. From 2009 to 2013, he led research in semantic search, graph data processing systems, and distributed query processing at Microsoft Research Asia. He had been a research staff member at IBM T. J. Watson Research Center from 2000 to 2009. He received his Ph.D. in Computer Science from the University of California, Los Angeles, in 2000. He has published more than 150 research papers in referred international journals and conference proceedings. He served as PC Chair of conferences such as SIGKDD'21 and is on the editorial board of journals such as IEEE Transactions of Knowledge and Data Engineering (TKDE) and Journal of Computer Science and Technology (JCST). He won the best paper award in ICDE 2015, the 10-year best paper award in ICDM 2013, and the best paper award of ER 2009.
Title: The Next Breakthrough in e-Commerce Search
Abstract: The effectiveness of Search has a substandtial impact on the revenue and growth of an e-commerce business. In this talk, I will discuss the current status and challenges of product search. In particular, I will highlight the enormous effort it takes to create a high-quality product search engine using classical information retrieval methods. Then, I will discuss how recent advances in NLP and deep learning, particularly the introduction of large pre-trained language models, may alter the status quo. While embedding-based retrieval has the potential to improve classical information retrieval methods, developing a machine learning-based, end-to-end system for general-purpose web search is still extremely difficult. Nevertheless, I will argue that product search for e-commerce may prove to be an area where deep learning can create the first disruption to classical information retrieval systems.
Hongtu Zhu
Professor of Statistics, University of North Carolina, Chapel Hill
Hongtu Zhu is a tenured professor of biostatistics, statistics, computer science, and genetics at University of North Carolina at Chapel Hill. He was DiDi Fellow and Chief Scientist of Statistics at DiDi Chuxing. He was Endowed Bao-Shan Jing Professorship in Diagnostic Imaging at MD Anderson Cancer Center. He chaired Departments of Statistical Cognitive Team and Feature Engineering with AI scientists and engineers on the development of innovative solutions for the world’s large ride-hailing platform at DiDi Chuxing. He is an internationally recognized expert in statistical learning, medical image analysis, precision medicine, biostatistics, artificial intelligence, and big data analytics. He has been an elected Fellow of American Statistical Association and Institute of Mathematical Statistics since 2011. He received an established investigator award from Cancer Prevention Research Institute of Texas in 2016 and received the INFORMS Daniel H. Wagner Prize for Excellence in Operations Research Practice in 2019. He has published more than 300 papers in top journals including Nature, Science, Nature Genetics, PNAS, AOS, and JRSSB, as well as 46 conference papers in top conferences including NeurIPS, AAAI, KDD, ICDM, MICCAI, and IPMI. He has served/is serving an editorial board member of premier international journals including Statistica Sinica, JRSSB, Annals of Statistics, and Journal of American Statistical Association.
Title: Advanced learning methods for two sided markets
Abstract: In this talk, we will introduce a general analytical framework for large scale data obtained from two-sided markets, especially ride-sourcing platforms like DiDi. This framework integrates classical methods including Experiment Design, Causal Inference and Reinforcement Learning, with modern machine learning methods, such as Graph Convolutional Models, Deep Learning, Transfer Learning and Generative Adversarial Network. We aim to develop fast and efficient approaches to address four major challenges for ride-sharing platform, ranging from demand-supply forecasting, demand-supply diagnosis, RL-based policy optimization, to A-B testing. Each challenge requires substantial methodological developments and inspires many researchers from both industry and academia to participate in this endeavor. All the research accomplishments presented in this talk were based on a series of joint works by a group of researchers at Didi Chuxing and my collaborators.