Original Thinking Lecture - Decision Making in Machine Learning

Our Original Thinking lecture series showcases some of the world-leading research from colleagues at Alliance MBS.

We all face and make decisions on an ongoing basis, whether at work or in our private lives. The vast majority of these decision involve trade-offs between multiple criteria, be it healthiness versus taste in our choice of breakfast cereal, cost versus energy efficiency in our choice of a new household utility, or risk versus expected return in our selection of a financial portfolio. Typically, there is considerable conflict between these criteria and, in the presence of such conflict, a single optimal solution may not exist. Taking a sound decision will then require the exploration of a set of alternative trade-offs, and the incorporation of additional preference information.

The same types of trade-offs exist in machine learning applications, where our models frequently have to strike a compromise between a variety of conflicting criteria. In this presentation, Julia will discuss the various origins of these criteria in a machine learning context. Using a number of examples from her own research, she will then highlight how multicriterion optimization can support us in exploring a range of alternative trade-off solutions for machine learning problems, supporting the analyst in identifying their preferred model.