The uncertainty of predictions from ML models is often just regarded as a bad thing - as a sign of model deficiencies, too little data, or lacking input features. In this talk, I will argue why knowing this uncertainty can be valuable. This especially applies if your ML model is used to minimize some kind of cost - either implicitly by a human or explicitly by an optimization algorithm.
The first step in realizing this value is going beyond point estimators towards predicting a probability distribution. In this talk, I will review how this can be done with
scikit-learn and related tools.
The second step is feeding these predictions into an optimization algorithm as probability distributions. As an example, I will present grocery store orders that are calculated from demand predictions. In extreme cases, including the uncertainty of the ML predictions turns out to be necessary for the optimization to be useful in practice at all.
Affiliation: Blue Yonder, a JDA company
2016 - present Data Scientist at Blue Yonder
- Development of ML models and autonomous replenishment systems
- Data Science support in supply chain projects
2015 Postdoctoral Researcher, University of Cologne
2010 - 2014 PhD in Computational Solid State Physics, RWTH Aachen
2003 - 2009 Studies of Physics, University of Ulm and University of New Mexico
visit the speaker at: Homepage