Optimization from Structured Samples —— An Effective Approach for Data-driven Optimization (Wei Chen)


Traditionally machine learning and optimizationare two different branches in computer science. They need to accomplish twodifferent types of tasks, and they are studied by two different sets of domainexperts. Machine learning is the task of extracting a model from the data,while optimization is to find the optimal solutions from the learned model. Inthe current era of big data and AI, however, such separation may hurt theend-to-end performance from data to optimization in unexpected ways --- arecent result shows a fundamental limitation that directly optimizing from datasamples is not achievable even when the separate model learning andmodel-driven optimization can be effectively executed. In this talk, I willintroduce an approach called optimization from structured samples (OPSS) totightly integrate learning and optimization by carefully utilizing thestructural information from the sample data to adjust the learning andoptimization algorithms. In particular, I will show how to overcome the abovelimitation when maximizing the (stochastic) coverage functions from structureddata samples even when a model cannot be accurately learned from the data. OPSSis an effective approach for the paradigm of data-driven optimization, and ithas applications in online advertising, influence maximization and otherdata-driven optimization tasks.


2021-06-18  14:00-14:30   


Wei Chen, Microsoft Research Asia


Guangdong Hotel Shanghai