During July 20th-25th, 2020, the 6thSummer School for Graduated Students of ZJU (ZJU-NUS International Summer School 2020 for Machine Learning and data-driven Decision Management) was held successfully through the ZOOM conference platform.
The summer school is organized by School of Management, International Research Center for Data Analytics and Management, Department of Data Science and Management Engineering, Zhejiang University, and jointly organized by the National University of Singapore. We invited 7 scholars in total, from both sides, the National University of Singapore and Zhejiang University. They deeply analyzed and generously shared related academic topics.
Prof. ZHOU Weihua (Associated Dean of School of Management, director of International Research Center of Data Analytics and Management of Zhejiang University) expressed his expectation for ZJU-NUS summer school combination in his welcome speech, and looked forward to the smooth development of international cooperation between both sides in the future in his conclusion speech. Prof. JIN Qingwei, Young Scholar PENG XiXian and Prof. WANG Mingzheng took turns being hosts during these six days.
1. Distributionally Robust Optimization
Professor Melvyn SIM from NUS, brought two wonderful reports on the theme of distributionally robust optimization. In the first lecture, Professor Sim starts with the concepts of uncertainty and risk preference and introduced the concepts of fuzzy sets and robust optimization. Combined with the classical newsboy model in operations research, and assuming that only part of the information about the requirements could be obtained, a new robust newsboy model could be obtained. Based on fuzzy set containing information demand in different ways, Prof. Sim introduced four conditions from easy to difficult, and gives the corresponding solving methods, which set a foundation for students’ learning. In the second lecture, Prof. Sim focused on the work of his team in solving a two-stage production-pricing problem using the distributionally robust optimization model, and their toolkit RSOME for robust stochastic optimization. He also suggested those students who want to learn more about robust optimization should start with the classic textbook, called Convex Optimization, and try to solve some problems using RSOME. During his lecture, Professor Sim left a deep impression on us with his careful and humorous explanation and penetrating reply.
2. Natural Language Understanding: Syntactic and Semantic Analyses
Dandan QIAO is an assistant professor from NUS. She brought the students a meaningful report from grammar to semantics, from vocabulary to sentence. First, she put forward the purpose of natural language processing succinctly: communication between machine and human being. Natural language processing is a challenging task due to the ambiguity of human languages among content, vocabulary and grammar. So she begins by detailing the semantic parsing method: Content-Free Grammars. Then she introduced many kinds of syntactic analysis methods, such as Constituency Parsing, Top Down Parsing, Bottom Up Parsing, Statistical Parsing, Probabilistic Context Free Grammar and so on. In the second part, she introduced Lexical Semantics, Vector Semantics, Word2vec and other semantic analysis methods. The last part of the lecture focused on how to express the meaning of a sentence clearly to a person in natural language processing. Her systematic and careful explanation won unanimous approval.
3. Smart Urban Transport and Logistics
Long HE is an assistant professor from NUS and he gave us a wonderful lecture all day along with the topic of smart city transportation and logistics. First, he briefly summarized the research direction and progress of smart city transportation and logistics in the field of OM/OR, mainly including technology innovation, business model innovation, last kilometer distribution, warehouse operation, medical scheduling, demand estimation, etc. Data-driven approaches have played a key role in many studies. After elaborating the ideas of raising and analyzing the problems, he mainly introduced two completed research works: the problem of taxi delivery in the taxi operation, as well as the problem of delivery order in the last kilometer. For problem 1, he used the method combining prediction and optimization to find the covariable and its relationship with the car demand in different situations through regression tree, so as to obtain a more refined fuzzy set. Then, he established and solved the distribution robust optimization model and obtained the corresponding dispatching strategy. For problem 2, he adopted a similar research framework, and used historical data to learn the dispatching time of drivers, and combined with the random variable of service time to establish the fractional robust optimization model and solve it. Professor He‘s report strategy of "teaching background in the morning and discussing methods in the afternoon" left enough time for participants to think, which was highly praised by audience.
4. Data Driven Platform Economics and Management
Professor Zike CAO from Zhejiang University, pointed out that platform business model and big data are two obvious trends at present in his report. The interaction between these two trends, that is, in the context of big data, is worthy of in-depth study of how the platform should formulate strategies. He shared a cross-sectional study of big data and platform strategies from three perspectives: 1. Merchant advertising and promotion: the impact of price discount on group purchasing, the effectiveness of making sales strategies based on shopping carts, and the information disclosure method of internet-celebrity-marketing on social media; 2. Platform governance: direct regulation and labeling reputation of merchants, whether the platform owner should enter the market, etc.; 3. User privacy and algorithm transparency: Privacy control and personalized advertising, etc. Last but not least, he encouraged participants to write an essay from the perspective of a reviewer, so as to find literature vacancy and new research direction, which greatly inspired students in the field of information systems research.
5. Topics in Dynamic Pricing
The price of a commodity often fluctuates due to supply and demand (e.g. seasonality, reference price), etc. An assistant professor from NUS, Zhenyu HU, provided models under two kinds of influence factors respectively, a dynamic pricing in reference price effect and consumers searching. He explained why price fluctuations occurs, and how to help merchants to obtain the optimal pricing strategy after understanding consumer behavior. In different price sensitivity, the demand function will also be different. He proposed that the profit function could be simplified by piecewise linear function, and showed the pricing strategy under different sensitivities: in the case of price reduction sensitivity, the income function is concave, and no price adjustment measures should be taken in the long run; However, under the condition of cyclic skimming pricing strategy, which has the nature of periodic promotion, the complexity of the problem will be greatly increased. In this case, the cyclic skimming pricing strategy is the optimal pricing scheme. In addition, merchant’s promotion in limited time will affect consumers search behavior. When the search is of little significance to consumers, merchants should give short and powerful discounts, but when search is of great significance to consumers, merchants should promote their products for a long time. When homogeneous products are promoted to a greater extent, price war is not a good strategy for merchants, so the intensity of discount should be controlled to extend the promotion time.
6. Panel Data Analysis with the Causal Inference
Nan CHEN is an assistant professor from NUS. He introduced the classical model in econometrics and mainly discussed the use of linear panel data model for causal inference. Professor Chen first introduced the classical linear model, least square estimation and its basic assumptions. He described the shortcomings of classical assumptions in the actual modeling process, and introduced two models of fixed effect model and random effect model, including the forms, basic assumptions, parameter estimation methods, model testing methods, etc. At the same time, the comparison between the two models is discussed, and the respective characteristics of the two models are used to guide how to choose the model when modeling practical problems. In addition, he explained the panel data exogeneity hypotheses, such as the concurrent exogeneity, and the challenges encountered in panel data modeling, such as bidirectional effects, unbalanced panels, and how to handle these issues based on the application domain.
7. Revenue Management under Discrete Choice Model
Professor Qingwei JIN of School of Management, Zhejiang University, introduced several types of discrete choice models, which are Locational Choice Model, the Multinomial Logit Model (MNL), Markov Chain Model and the Random Utility Model (RUM). The MNL Model is a special circumstance of Markov Chain Model, and Markov Chain Model and Locational Choice Model are special cases of the RUM. He then introduced the connection between these models and the Binary Choice Forest. He illustrated the specific application of MNL Model and Markov Chain Model in combinatorial optimization problems, as well as the application of MNL Model in the joint optimization of pricing and product combination problem, as well as comparing the performance of MNL Model with machine learning algorithm in product recommendation. The results showed that although the MNL model is inferior to the machine learning algorithm in terms of prediction accuracy, it brings much higher expected benefits than the machine learning algorithm. Finally, he explained how to make pricing and combinatorial optimization when there is reference effect.
On the last day of the summer school, we held a panel discussion, the speakers answered questions online for the students. One of the most concerned questions is how to choose their own research direction. They shared that the choice of research direction is not an overnight process. It is necessary to combine self-interests and extensive literature reading. It is inevitable to make trial and error in the early stage, but don‘t be afraid of making mistakes.
This summer school adopts the online teaching method, the number of registered students and the total number of participants in a week are all over 1,000, and the participants enjoy the academic feast together for a week. With the excellent communication platform, we had a further understanding of the specific issues in machine learning, operation optimization, supply chain, revenue management and other aspects of studies, and many participants took the opportunity to have in-depth discussion with speakers. Meanwhile, speakers are also looking forward to further collaborative projects and opportunities to jointly tutor students in the future.
