A semi-supervised active learning approach for block-status classification
Atul Rawal, James McCoy, Andrew Duvall, Elvis Martinez · Papers in Applied Geography · 2025
The Census Bureau, as a part of its decennial census must maintain and update all the addresses present within the United States and its territories. These addresses help formulate policies and allocate valuable resources from the federal government. For the 2020 Census, in-office staff manually canvased address coverage in every block. While this process was effective, it also brought about challenges associated with cost and time. We present a robust machine learning solution to improve both data labeling and classification of parcel data to enable new data-driven insight while reducing costs and effort for data assessment. We utilize an active-learning scheme to make accurate/precise classifications using the <1% labeled blocks out of the 8,000,000+ blocks within the country. We utilized multiple machine learning models to make predictions on unlabeled data by training the model on the smaller set of labeled data. Predictions from all the models are then compared to pinpoint the blocks where there is a mismatch between the different models. The mismatched blocks are then forwarded to the human labelers to make a final prediction. We also discuss the different challenges associated with working on real-world data at this scale such as class-imbalance and data completeness, integrity.