United StatesPublic sector2021 – ongoing

U.S. Census Bureau

Commodity Flow Survey machine-learning autocoding

Census reported using machine learning to correct or fill codes for approximately 200,000 shipment records

WorkingCompany's own claim

Problem

Free-text shipment descriptions needed product codes for the Commodity Flow Survey. Looking up codes created work for analysts and left records requiring correction.

Approach

Census implemented a machine-learning autocoder for product descriptions. It used the model to impute or relabel records from the 2017 survey. An analyst search tool ranked likely codes to narrow manual lookup.

Results

approximately 200,0002017 survey shipment records imputed or relabeled, per CensusCompany's own claim[1]

10candidate codes shown to analysts from 514 possibilities, per CensusCompany's own claim[1]

Takeaway

The same classification model supported direct record coding and analyst review. Projected respondent savings were separate from completed record processing.

Sources

Modernizing CFS Collection and Operations with Machine Learning
American Statistical Association15 Jun 2021Research paper or report

Source summaries

[1]American Statistical Association15 Jun 2021

A Census presenter described implemented autocoding and analyst search tools. The abstract reported completed record corrections and candidate-code counts, while respondent cost savings remained a forecast.

Full details are in the original sources.

Added 1 Nov 2026 · Updated 1 Nov 2026