Business Context

OpenAI’s billing risk team trains a fraud classifier daily to score card payment attempts and account-level abuse signals. The current training job takes 6+ hours, which delays model refreshes and slows incident response when fraud patterns shift.

Dataset

You are given a tabular binary classification dataset built from 90 days of payment attempts and account activity.

• Size: 12.4M payment attempts, 55 features
• Target: is_fraud — chargeback or confirmed payment abuse within 14 days
• Class balance: 1.6% positive, 98.4% negative
• Missing data: 8% missing in issuer/card metadata, 3% missing in account aggregates for new users

Success Criteria

A solution is good enough if it reduces end-to-end training time by at least 60% while keeping PR-AUC within 2% relative of the current production model and preserving recall at 5% alert rate.

Constraints

• Daily retraining budget is limited to a single 16-core CPU machine; GPU is not guaranteed.
• Batch scoring latency is not critical, but training turnaround is.
• The risk team needs feature importance and a reproducible pipeline.
• Avoid leakage from future chargeback labels or post-authorization events.

Deliverables

1. Propose a training strategy that optimizes time-to-result, not just final offline score.
2. Build a baseline and at least one accelerated training approach.
3. Justify sampling, feature reduction, and hyperparameter search choices.
4. Evaluate quality vs runtime tradeoffs using a time-based validation setup.
5. Recommend what should ship to production and how often it should retrain.