Machine learning for official statistics: UNECE help statistical organisations harness the power of machine learning
The new UNECE publication Machine Learning for Official Statistics helps national and international statistical organizations harness the power of machine learning to modernize the production of official statistics.
Machine learning in the context of modernisation for statistical organizations
Statistical organizations produce crucial indicators that portray various aspects of the economy and society that we live in. These include measures such as the gross domestic product, the inflation rate, the population growth, and the unemployment rate on which government and business alike depend when making important decisions. While these may be just a few digit numbers from the end-user side, statistical organizations employ a series of carefully designed and executed processes to distil this key information from the vast amount of raw data.
With increasing challenges arising from new data sources, technological developments and competitions with private companies, statistical organizations have been striving to modernise every part of this production process to provide more relevant and detailed official statistics in a more timely and accessible manner. They utilise the computer-assisted interview and web scraping tool to collect data more efficiently, and build infrastructure for data and IT tools to manage them across the organizations more easily.
Yet, one area that is difficult to modernise is the processes that require “human-like” decision-making, such as reading a textual description to assign a matching classification code or looking at the image to identify what it represents. Traditionally, this has been done either manually or through a complex rule-based system, both of which are costly, time-consuming and hard to manage. This is particularly daunting when statistical organizations try to use big data sources (e.g., price information web-scrapped from online stores) as the cost of resources needed to process such a large amount of data in the traditional manual way is simply too prohibitive.
Machines learning holds great potential for statistical organizations
The recent developments in machine learning technique are pushing the boundary of tasks considered for humans and machines - machines can now draw a painting in the style of an old master and write an article just like humans.
How does this technology work? In one of the most popular approaches called “supervised learning”, machines are first trained on the data that humans labelled, for example, images labelled as “urban” or “rural”. With this data, they figure out patterns associated with labels by adaptively improving their internal logic that maps from the input (image) to output (label). In this way, machines can determine whether an area shown in an image is urban or rural without us providing all possible rules explicitly.
As the machine learning technique can carry out tasks that we used to solely rely on manual works, it holds a great potential to increase the efficiency of statistical organizations, just like the use of machinery powered by steam engines made a huge leap in the productivity in the manufacturer industry few centuries ago. Also, their capability to process various types of data such as text, image and video offers statistical organizations to take advantage of new data sources to produce new statistics that could meet the evolving needs of society.
Challenges in using machine learning for official statistics
Like with any innovation, however, the journey of integrating machine learning in the organization abounds with challenges and setbacks. The technology itself is still relatively new and requires a different skill set that many statistical organizations do not possess; hence it needs to be built inside or acquired from outside.
The real difficulty, however, starts when the machine learning solution needs to move to production, meaning that it is connected to existing processes seamlessly and used for the regular business, beyond an “experiment” stage. Unfortunately, even after successful pilot studies, many machine learning solutions end up being left on the shelf. The difficulty is experienced widely across sectors and domains. It is said that over 80% of machine learning projects never make it to production. Moving machine learning into production requires changes in infrastructure, culture, organizational structure or business processes, none of which is a small task with a lasting effect.
UNECE supports statistical organizations in advancing the use of machine learning
Based on the two international initiatives, the UNECE High-Level Group for the Modernisation of Official Statistics (HLG-MOS) Machine Learning Project (2019-20) and the United Kingdom Office of National Statistics (ONS) – UNECE Machine Learning Group 2021, the publication Machine Learning for Official Statistics aims to help statistical organizations navigate the difficult journey of advancing the use of this new technology. It presents the practical applications of machine learning in three working areas within statistical organizations and discusses their value-added, challenges and lessons learned. The publication also includes a quality framework that could help guide the choice of methods, demonstrates key steps for moving machine learning from the experimental stage to the production stage, and key messages to facilitate the use of machine learning in the statistical organizations.
The machine learning field is fast evolving with new methods, platforms and approaches coming out every month. To keep up with the pace of change and avoid duplication of efforts, there is a great need for knowledge sharing and collaboration within the official statistics community. UNECE continues its engagement in the international initiative this year, through Machine Learning Group 2022 with the ONS, to support statistical organizations to harness the power of machine learning.
