Abstract
In official statistics, machine learning applications continue to increase, and conventional error metrics are frequently used to assess model performance. Despite this, these metrics may not capture all of the errors that can adversely impact the validity of machine learning models. An integrated framework for addressing and assessing the multiple sources of machine learning errors is provided by the Total Machine Learning Error (TMLE) model. Developed from the well-established Total Survey Error (TSE) model, the model emphasizes the importance of analyzing multiple error components, including over- and under-coverage, sampling error, errors caused by model assumptions, and measurement errors. In this colloquium, I will discuss several errors associated with the TMLE model. In official statistics, these errors play a critical role in the applicability of machine learning.
 

Marco Puts currently works as a Methodologist at Statistics Netherlands and focuses on Machine learning Methodology and perceptual aspects of visualization.
 

The RSS Manchester local group (contact antonia.marsden@manchester.ac.uk).