HaiAI123

Curated Global AI Tools Directory

Industry update
25
heat index
Flat

Developer shows preprocessing before split inflates offline score by 4.2 points

11/14/2026 — 11/14, 10:18·1 sources (all)·1 reports (all)

Key takeaways

  • Using an imbalanced digits dataset, the author ran a controlled comparison: SMOTE on the full set before splitting gave 0.968 offline macro…

Background & analysis

On November 14, 2026, a CNBlogs post published as the seventh entry in the "AI Engineer Pitfall Notes" series examined how doing preprocessing before data splitting inflates offline evaluation scores. The author built an imbalanced dataset from the digits handwritten-digit data, with 175 images each for majority classes 0 through 6 and 12 images each for minority classes 7 through 9, then applied SMOTE oversampling to balance the minority classes. Both arms of the controlled experiment ran the same network with the same random seed and shared an identical base training set; the only difference was that the leakage arm additionally mixed in 96 synthetic samples generated by interpolating test-set samples with their nearest neighbors. This design reproduces the behavior of oversampling on the full dataset before splitting: when SMOTE interpolates over all data, its parent samples include those later assigned to the test set, so synthetic samples become close relatives of test samples and end up in the training set.

Under the correct approach of splitting first and oversampling only within the training set, the offline macro F1 reached 0.926 and performance on new online data was 0.786. Under the leaked approach of oversampling the full dataset before splitting, the offline macro F1 was 0.968 and online performance on new data was 0.819, meaning the leak inflated the offline score by 4.2 points. The author also reported honestly that normalization leakage showed no gap on this dataset. Based on these findings, the author recommends treating the split as the first line of the workflow, placing preprocessing inside the Pipeline, and running a final holdout evaluation before deployment, with the companion experiment script released on the author's official account.

AI-generated from 1 reports · updated 3 days ago

updream
24-hour heatHeat index 25
Just charted — trend is building up
24 hours agonow

Reports on this story headlines open the original

Nov 14
  1. A controlled experiment on an imbalanced digits dataset shows that running SMOTE before the train/test split inflates offline macro-F1 by 4.2 points: 0.968 offline versus 0.819 on new data, against 0.926 and 0.786 for the correct order. The author also reports a null result for scaler leakage on this data and argues the split must always come first, with preprocessing inside a Pipeline.

    博客园AI score 64

Other stories people are talking about

How is heat calculated?About the method

Heat counts how many independent sources covered a story in the last 48 hours: one source counts once no matter how many posts it published, decaying with a 24-hour half-life. What ranks first is what many people are talking about.

This page aggregates public feeds. Headlines and summaries are machine-organized and remain the property of the original authors; verify important facts at the source. If you believe a headline or summary infringes your rights, email the contact address in the footer with the page URL and basis of your claim, and we will remove or replace it after review.

Surge
Discussion rising fast
New
First report within 6 hours
Rising
Still gathering discussion

Back to AI industry updates →