Optimization from Structured Samples —— An Effective Approach for Data-driven Optimization (Wei Chen)
Abstract
Traditionally machine learning and optimization
are two different branches in computer science. They need to accomplish two
different types of tasks, and they are studied by two different sets of domain
experts. Machine learning is the task of extracting a model from the data,
while optimization is to find the optimal solutions from the learned model. In
the current era of big data and AI, however, such separation may hurt the
end-to-end performance from data to optimization in unexpected ways --- a
recent result shows a fundamental limitation that directly optimizing from data
samples is not achievable even when the separate model learning and
model-driven optimization can be effectively executed. In this talk, I will
introduce an approach called optimization from structured samples (OPSS) to
tightly integrate learning and optimization by carefully utilizing the
structural information from the sample data to adjust the learning and
optimization algorithms. In particular, I will show how to overcome the above
limitation when maximizing the (stochastic) coverage functions from structured
data samples even when a model cannot be accurately learned from the data. OPSS
is an effective approach for the paradigm of data-driven optimization, and it
has applications in online advertising, influence maximization and other
data-driven optimization tasks.
Time
2021-06-18 14:00-14:30
Speaker
Wei Chen, Microsoft Research Asia
Room
Guangdong Hotel Shanghai
