Events & Seminars

Emission from diesel trucks such as Nitrogen Oxides causes severe air pollution. However, direct regulation on trucking companies for their use of diesel trucks typically falls out of the jurisdiction of local governments. A legislative alternative is to regulate other sectors that prompt diesel truck usage in the local region, called indirect emission sources. The first of such regulations is Southern California's Rule 2305, the Warehouse Indirect Source Rule (ISR). Passed in May 2021, the ISR holds local warehouses responsible for the diesel truck trips to their facilities through a mitigation fee. The goal of the ISR is to incentivize warehouses to hire electric semi-trucks to improve air quality and thus public health. Motivated by this new policy, we explore the environmental impact of the ISR and the industry burden that it introduces, compared to a hypothetical direct source rule (DSR) that regulates trucking companies. We find that ISR can indeed lead to higher adoption of electric semi-trucks than DSR, especially when the mitigation fees for diesel truck trips are small. However, using the mitigation fee collected from warehouses to subsidize trucking companies’ electric semi-truck investments, a current practice of the ISR, can backfire and reduce industry adoption of electric semi-trucks. Interestingly, depending on the distribution of truck trips' distances, a higher mitigation fee for using diesel trucks can also lead to lower adoption of electric semi-trucks. Finally, we explore the practical implications of the ISR using real warehouse data from Southern California.

The undesirable but inevitable consequence of running promotions is that consumers can be trained to time their purchases strategically. In this paper, we study randomized promotions, where the firm randomly offers discounts over time, as an alternative strategy of intertemporal price discrimination. Specifically, we consider a base model where a monopolist sells a single product to a market with a constant stream of two market segments. The segments are heterogeneous in both their product valuations and patience levels. The firm pre-commits to a price distribution, and in each period, a price is randomly drawn from the chosen distribution. We characterize the optimal price distribution as a randomized promotion policy and show that it serves as an intertemporal price discrimination mechanism such that high-valuation customers would purchase immediately at a regular price upon arrival and low-valuation customers would wait for a random promotion. Compared against the optimal cyclic pricing policy, which is optimal within the strategy space of all deterministic pricing policies, the optimal randomized pricing policy beats the optimal cyclic pricing policy if low-valuation customers are sufficiently patient and the absolute discrepancy between high and low customer valuations is large enough. We extend the model in two directions. We first consider Markovian pricing policies where prices are allowed to be intertemporally correlated in a Markovian fashion. This additional maneuver allows the firm to reap an even higher profit when low-valuation customers are sufficiently patient, by avoiding consecutive promotions but on average running the promotion more frequently with a smaller discount size. We then consider a model with multiple customer segments, and show that a two-point price distribution remains optimal and our conclusion from the two-segment base model still holds under certain conditions that are adopted in the literature. Our results imply that the firm may want to deliberately randomize promotions in the presence of forward-looking customers.

Causal decision making (CDM) at scale has become a routine part of business, and increasingly, CDM is based on machine learning. Businesses algorithmically target offers, incentives, and recommendations to affect consumer behavior. Recently, we have seen an acceleration of research related to CDM and causal effect estimation (CEE) using machine-learned models. This article highlights an important perspective: CDM is not the same as CEE, and counterintuitively, accurate CEE is not necessary for accurate CDM. Technically, the estimand of interest is different, and this has important implications both for modeling and for the use of statistical models for CDM. In this talk, I will highlight three implications. (1) We should carefully consider the objective function of the causal machine learning, and if possible, optimize for accurate “treatment assignment” rather than for accurate effect-size estimation. (2) Confounding affects CDM and CEE differently. The upshot here is that for supporting CDM it may be just as good or even better to learn with confounded data as with unconfounded data. (3) Causal statistical modeling may not be necessary at all to support CDM because a proxy target for statistical modeling might do as well or better. This third observation helps to explain at least one broad common CDM practice that seems “wrong” at first blush—the widespread use of noncausal models for targeting interventions. The last two implications are particularly important in practice, as acquiring (unconfounded) data on both “sides” of the counterfactual for modeling can be quite costly and often impracticable. These observations also open substantial research ground that I will discuss at the end of the talk.

Personalized intervention management in healthcare has received a rapidly growing interest in the big-data era yet still is a burgeoning field. A key challenge for personalization in healthcare is data scarcity. This small sample issue makes standard learning methods hard to learn the right policy and/or suffer from large variances. In this research, we extend the data-pooling technique from the bandit setting to the reinforcement learning (RL) context. RL models explicitly account for future cost/reward and are more suitable for healthcare management problems. We develop a novel data-pooling estimator in the RL context, and establish theoretical performance guarantee for RL with data-pooling. We demonstrate its empirical success on a real hospital dataset with an application to reduce 30-day hospital readmission rate. This is a joint work with Xinyun Chen and Xiuwen Wang from CUHK Shenzhen.

We study the problem of community detection in multi-layer networks, where pairs of nodes can be related in multiple modalities. We introduce a general framework, i.e., mixture multi-layer stochastic block model (MMSBM), which includes many earlier models as special cases. We propose a tensor-based algorithm (TWIST) to reveal both global/local memberships of nodes, and memberships of layers. We show that the TWIST procedure can accurately detect the communities with small misclassification error as the number of nodes and/or the number of layers increases. Numerical studies confirm our theoretical findings. To our best knowledge, this is the first systematic study on the mixture multi-layer networks using tensor decomposition. The method is applied to two real datasets: worldwide trading networks and malaria parasite genes networks, yielding new and interesting findings.

Based on joint work with Bing-Yi Jing, Ting Li and Zhongyuan Ly.

We study the matching of jobs to workers in a queue, e.g. a ridesharing platform dispatching drivers to pick up riders at an airport. Under FIFO dispatching, the heterogeneity in trip earnings incentivizes drivers to cherry-pick, increasing riders' waiting time for a match and resulting in a loss of efficiency and reliability. We first present the direct FIFO mechanism, which offers lower-earning trips to drivers further down the queue. The option to skip the rest of the line incentivizes drivers to accept all dispatches, but the mechanism would be considered unfair since drivers closer to the head of the queue may have lower priority for trips to certain destinations. To avoid the use of unfair dispatch rules, we introduce a family of randomized FIFO mechanisms, which send declined trips gradually down the queue in a randomized manner. We prove that a randomized FIFO mechanism achieves the first best throughput and the second best revenue in equilibrium. Extensive counterfactual simulations using data from the City of Chicago demonstrate substantial improvements of revenue and throughput, highlighting the effectiveness of using waiting times to align incentives and reduce the variability in driver earnings.

Job postings provide unique insights about the demand for skills, tasks, and occupations. Using the full text of data from millions of online job postings, we train and evaluate a natural language processing (NLP) model with over 100 million parameters to classify job postings' occupation labels and salaries. To derive additional insights from the model, we develop a method of injecting deliberately constructed text snippets reflecting occupational content into postings. We apply this text injection technique to understand the returns to several information technology skills including machine learning itself. We further extract measurements of the topology of the labor market, building a ``jobspace'' using the relationships learned in the text structure. Our measurements of the jobspace imply expansion of the types of work available in the U.S. labor market from 2010 to 2019. We also demonstrate that this technique can be used to construct indices of occupational technology exposure with an application to remote work. Moreover, our analysis shows that data-driven hierarchical taxonomies can be constructed from job postings to augment existing occupational taxonomies like the SOC (Standard Occupational Classification) system.

The U.S. Securities and Exchange Commission (SEC) requires firms to notify investors in an SEC filing of a data breach if it constitutes a material event. Importantly, the determination of materiality lies with executives, which has resulted in firms failing to disclose breaches to the SEC or purposely delaying notifications. We draw from the behavioral theory of the firm and executive compensation literature to develop predictions about the influence of IT and non-IT executives’ compensation on firms’ SEC data breach notification delays. Given the possibility of competing priorities and goals of the two executive groups, we argue that increased IT executive compensation leads to fewer delays, whereas increased non-IT executive compensation has the opposite effect. Because corporate boards of directors have oversight and advise on firms’ cybersecurity matters, we argue that the cybersecurity intensity of the firm’s board (i.e., social ties to breached firms) moderates the relationships between IT and non-IT executive compensation and notification delays. To test our hypotheses, we constructed a panel dataset from public sources and performed a series of econometric analyses. Our results suggest that the influence of executive compensation on notification delays differs for IT and non-IT executives in the manner hypothesized. However, for both types of executives, the moderating influence of the board’s cybersecurity intensity works to increase notification delays. Counter to the conventional view that increased cybersecurity experience on the board benefits timely data breach notification, our findings suggest that greater board experience results in delays of timely communications about data breaches via 8-K filings.

Operating under both supply-side and demand-side uncertainties, a mobile-promotion platform conducts advertising campaigns for individual advertisers. Campaigns arrive dynamically over time, which is divided into seasons; each campaign requires the platform to deliver a target number of mobile impressions from a desired set of locations over a desired time interval. The platform fulfills these campaigns by procuring impressions from publishers, who supply advertising space on apps, via real-time bidding on ad exchanges. Each location is characterized by its win curve, i.e., the relationship between the bid price and the probability of winning an impression at that bid. The win curves at the various locations of interest are initially unknown to the platform, and it learns them on the fly based on the bids it places to win impressions and the realized outcomes. Each acquired impression is allocated to one of the ongoing campaigns. The platform's objective is to minimize its total cost (the amount spent in procuring impressions and the penalty incurred due to unmet targets of the campaigns) over the time horizon of interest.

Our main result is a bidding and allocation policy for this problem. We show that our policy is the best possible (asymptotically tight) for the problem using the notion of regret under a policy, namely the difference between the expected total cost under that policy and the optimal cost for the clairvoyant problem (i.e., one in which the platform has full information about the win curves at all the locations in advance): The regret under any policy is , where is the number of seasons, and that under our policy is . We demonstrate the performance of our policy through numerical experiments on a test bed of instances whose input parameters are based on our observations at a real-world mobile-promotion platform.

Perception regarding a focal firm's corporate social responsibility (CSR) depends not only on itself but also on its known suppliers. This paper provides the first empirical evidence linking CSR and supply chain information disclosure together. We uncover robust evidence that listed firms voluntarily disclose environmentally responsible suppliers while selectively not disclosing" bad" ones, effectively greenwashing their supply chain image. This selective disclosure of green suppliers is prevalent among listed firms across the world. Such corporate behavior is increasing in public awareness of climate change, decreasing in regulations on CSR information transparency. It is more salient for firms who face higher competition or care more about their brand awareness, and for firms that are more profit-driven or held more by institutional investors. Firms that greenwash supply chains observe an increase in sales and valuation, suggesting that consumers and investors do not fully take greenwashing of listed firms into account.