What is it about?

National statistical offices are increasingly using big data and machine learning to improve the production of official statistics. However, building an accurate model is only part of the challenge. Statistical organizations also need reliable ways to deploy, monitor, maintain, and govern machine-learning systems in production. This study reviews the literature on big data, data science, and machine learning in official statistics using Scopus and Web of Science. A total of 178 research articles were identified. The review shows that much of the existing research focuses on alternative data sources and machine-learning techniques, while considerably less attention has been given to what happens after model development. Important gaps remain in machine learning operations (MLOps), production deployment, data quality assurance, governance, and the adoption of feature stores.

Featured Image

Why is it important?

Official statistics must meet particularly high standards of reliability, reproducibility, transparency, and data quality. Machine-learning models that work well in experimental settings therefore cannot simply be transferred into production without appropriate operational and governance processes. Our review highlights an important imbalance in the literature: substantial research has examined how machine learning and alternative data can improve statistical production, but much less attention has been paid to deploying and maintaining these systems reliably over time. Greater attention to MLOps, feature management, data quality, monitoring, and governance can help national statistical offices move from isolated machine-learning experiments toward scalable and dependable production systems.

Perspectives

The future of machine learning in official statistics depends not only on developing better predictive models, but also on building the infrastructure and governance needed to operate them reliably. Our review suggests that the post-training stages of the machine-learning lifecycle remain comparatively underexplored. Questions surrounding model deployment, monitoring, data quality, governance, and feature management become increasingly important as statistical organizations move from experimentation toward operational use. MLOps and feature stores provide promising mechanisms for making these processes more reproducible, scalable, and manageable. Future research should therefore look beyond model accuracy and examine how machine-learning systems can be maintained as trustworthy components of official statistical production.

Prof. Afshin Ashofteh
Universidade Nova de Lisboa

Read the Original

This page is a summary of: A Review of Big Data and Machine Learning Operations in Official Statistics: MLOps and Feature Store Adoption, July 2024, Institute of Electrical & Electronics Engineers (IEEE),
DOI: 10.1109/compsac61105.2024.00101.
You can read the full text:

Read

Contributors

The following have contributed to this page