Colloquium IS (November 3, 2017) – Interplay between learning and optimization

Dear all,

We are pleased to invite you to our next Colloquium IS that will take place on Friday,

November 3, 2017, 12:30 – 13:30 (Paviljoen K.16).

Speaker: Yingqian Zhang

Title: Interplay between learning and optimization

Abstract:

There are increasing interests in combining machine learning and optimization. In this talk, I will introduce two ongoing work in this research line.

In most existing approaches of using data to solve optimization problems, predictive (machine learning) models serve as decision variables, input parameters, or solution evaluation functions. In our work, we show how to use the internal structure of predictive models in optimization process, and demonstrate how the proposed approach helps to find better solutions.

Predictive models such as decision trees are typically built using sub-optimal algorithms, which often aim at optimizing loss functions (i.e., accuracy). These algorithms are not flexible when a different learning objective rather than accuracy is desired. We propose to transfer the decision tree learning problem to a mathematical optimization problem. In this way, different learning objectives, such as minimizing discrimination or false positive errors, can be easily specified for constructing optimal predictive models.

I will use online auction as an example to illustrate both approaches. 

 

Comments are closed.