Events & Seminars

Cold start describes a commonly recognized challenge in online advertising platforms: With limited data, the machine learning system cannot accurately estimate the click-through rates (CTR) nor the conversion rates (CVR) of new ads and in turn cannot efficiently price these new ads or match them with platform users. Unsuccessful cold start of new ads will prompt advertisers to leave the platform and decrease the thickness of the ad marketplace. To address the cold start issue for online advertising platforms, we build a data-driven optimization model that captures the essential trade-off between short-term revenue and long-term market thickness of advertisement. Based on duality theory and bandit algorithms, we develop the Shadow Bidding with Learning (SBL) algorithm with a provable regret upper bound of O(T^{2/3}K^{1/3}(log T)^{1/3}d^{1/2}), where K is the number of ads and d is the effective dimension of the underlying machine learning oracle for predicting CTR and CVR. Furthermore, our proposed algorithm can be straightforwardly implemented in practice with minimal adjustments to a real online advertising system. To demonstrate the practicality of our cold start algorithm, we collaborate with a large-scale online video sharing platform to implement the algorithm online. In this context, the traditional single-sided experiment would result in substantially biased estimates. Therefore, we conduct a novel two-sided randomized field experiment and devise unbiased estimates to examine the effectiveness of the SBL algorithm. Our experimental results show that the proposed algorithm could substantially increase the cold start success rate by 61.62% while only compromising the short-term revenue by 0.717%. Our new algorithm has also boosted the overall market thickness by 3.13% and the long-term life-time advertising revenue by at least 11.16%. Our study bridges the gap between the bandit algorithm theory and the ads cold start practice, and highlights the significant value of well-designed cold start algorithms for online advertising platforms.

We study the fundamental stochastic newsvendor model that considers both demand and yield randomness. Although partial statistical information and empirical data are often accessible, it is usually difficult in practice to describe precisely the joint demand and yield distribution. We combat the issue of distributional ambiguity by taking a data-driven distributionally robust optimization approach. We adopt the minimax regret decision criterion to assess the optimal order quantity that minimizes the worst-case regret across all hedged distributions. Then we present several properties about the minimax regret model, including optimality condition, regret bound, and worst-case distribution, and we show that the optimal order quantity can be determined via an efficient golden section search. Finally, we present numerical comparisons of our data-driven minimax regret model with data-driven models based on Hurwicz decision criteria and with a minimax regret model based on partial statistical information on moments.

This is a teaching demonstration of a class on the topic of hypothesis testing for a single population mean. The class will start with examples of real-life applications to stimulate the audience’s interest in the topic. Fundamentals of the hypothesis test for a single population mean will be addressed. The class will be concluded with examples of hypothesis testing with both continuous data and dichotomous data.

The booming of third-party logistics (3PL) changes the cost structure of an online retailer in the order fulfillment process. The online retailer pays a fixed amount of order arrangement fee to the 3PL to outsource the order fulfillment service for each service request. We study the problem of when an online retailer should send the service request. The trade-off is between the order arrangement fee and the order holding cost. We model the problem as a Markov Decision Process (MDP). By reducing the MDP to a sequence of single-dimensional counterparts, we analytically characterize the optimal order-holding policy. To calculate the policy, we apply a consumer sequential choice model to characterize the transition probabilities, which captures the heterogeneity across different orders and admits a personalized order-holding policy. We further get the closed form of the personalized order-holding policy and provide a piecewise linear approximation of the policy. Extensive numerical tests based on the data set from the 2020 MSOM Data-Driven Research Challenge show that (1) The gap of piecewise linear approximation is as small as 1; (2) The proposed policy achieves a considerable cost reduction compared to two benchmarks in the literature, with an average 30.12% and 14.01% cost reduction for enterprise users in all instances compared with two other widely used policies in the literature, respectively.

With food delivery services, customers can hire delivery workers to pick up food on their behalf. To investigate the long-term impact of food delivery services on the restaurant industry, we model a restaurant serving food to customers as a stylized single-server queue with two streams of customers. One stream consists of tech-savvy customers who have access to a food delivery service platform. The other stream consists of traditional customers who are not able to use a food delivery service and only walk in by themselves. We study a Stackelberg game, in which the restaurant first sets the food price; the food delivery platform then sets the delivery fee; and, last, rational customers decide whether to walk in, balk, or use a food delivery service if they have access to one. We show that the food delivery platform does not necessarily increase demand for the restaurant but may just change the composition of customers, as the segment of tech-savvy customers grows. Hence, paying the platform for bringing in customers may hurt the restaurant's profitability. We demonstrate that a one-way revenue-sharing contract with a price ceiling or a two-way revenue-sharing contract can coordinate the system and create a win-win. Furthermore, under conditions of no coordination between the restaurant and the platform, we show, somewhat surprisingly, that more customers having access to a food delivery service may hurt the platform itself and the society, when the food delivery service is sufficiently convenient and the pool of delivery workers is large enough. This is because the restaurant can become a delivery-only kitchen and raise its food price by focusing on food-delivery customers only, leaving little surplus to the platform. This implies that limiting the number of delivery workers can provide a simple yet effective means for the platform to improve its own profit while benefiting the social welfare.

Google search indices can be useful for measuring time-varying cross-regional public interests for which survey data are extremely rare. However, there is a practical difficulty with generating longitudinal Google Trends. Google Trends provides normalized counts from zero to 100 instead of absolute counts, thereby placing its cross-sectional indices across different times on different scales. Thus, merely pooling cross- sectional data fails to create desirable longitudinal data. To resolve this problem, we develop a method for rescaling Google Trends indices to build longitudinal data. We illustrate this method with applications to the issues of employment and the coronavirus. This new tool opens the door to using Google searches merged with various kinds of time-series cross-sectional data, which has not been possible.