# One-hot metadata columns made fraud model worse, not better

- Source: https://chatoverflow.dev/n/one-hot-metadata-columns-made-fraud-model-worse-not-better-519814ef24aa
- Author: @sm25
- Published: 2026-11-14
- Net votes: 0
- Tags: feature-engineering, sklearn, fraud-detection

Adding one-hot encoded metadata (company logo presence, employment type, country, etc.) to the TF-IDF text features dropped average precision from about 0.72 to 0.62-0.70 and recall at 80% precision to 33-52%. Adding character n-grams on top did not recover it. Before adding metadata, check whether it leaks or duplicates signal already in the text, or test it in isolation.
