Business Context
You’re on the Risk ML team at MercuryPay, a fintech processing $18B/year in card-not-present payments across North America and the EU. Fraud losses are rising, and the company is rolling out a new real-time fraud model that must score transactions in <50 ms p95. The current feature store has grown organically and now produces thousands of features per transaction; training is slow, inference is expensive, and model performance is unstable due to multicollinearity and drift.
Your task is to design and justify a dimensionality reduction strategy that improves training efficiency and generalization while meeting latency and compliance constraints.
Dataset
MercuryPay logs one row per transaction with engineered features derived from user behavior, merchant history, device fingerprinting, and network signals.
| Feature Group | Count | Examples | Notes |
|---|
| Numeric aggregates | 1,850 | avg_spend_7d, txn_count_1h, chargeback_rate_90d | Heavy-tailed, many correlated |
| Categorical | 120 | merchant_mcc, country, card_bin, device_os | High-cardinality (BIN ~30K) |
| Sparse text-derived | 600 | hashed_email_ngram_, merchant_name_char_3gram_ | Very sparse |
| Graph/network | 90 | shared_device_degree, merchant_cluster_id | Some leakage risk if built incorrectly |
| Temporal | 60 | hour_of_day, days_since_first_seen | Cyclical patterns |
- Size: ~48M transactions over 6 months; you can assume a sampled training set of 8M rows fits in a single training job.
- Target:
is_fraud (chargeback-confirmed within 60 days).
- Class balance: 0.42% fraud (highly imbalanced).
- Missing data: ~12% missing in device signals (privacy opt-outs), ~4% missing in merchant enrichment.
Success Criteria
- Model quality: Improve PR-AUC by +0.02 absolute over a strong baseline (regularized logistic regression on all features).
- Operational: Reduce feature dimensionality by ≥80% while keeping p95 inference latency < 50 ms.
- Stability: Reduce week-to-week metric variance (PR-AUC std dev) by ≥20% on rolling weekly backtests.
- Interpretability & compliance: Provide a clear explanation of what information is preserved/removed; avoid transformations that make it impossible to audit key risk drivers in the EU.
Constraints
- No target leakage: any reduction must be fit only on training folds/time windows.
- Time-aware evaluation: random splits are not allowed; fraud patterns drift.
- Production: the transformation must be reproducible, versioned, and fast at inference.
- Feature types: mix of dense numeric, high-cardinality categorical, and sparse hashed features.
Deliverables (what you would present in an interview)
- Propose at least three dimensionality reduction techniques (e.g., filter-based, embedded, and projection-based) and explain when you’d choose each for this fraud setting.
- Provide a concrete plan for handling sparse + dense features (e.g., separate pipelines, hybrid approaches).
- Define a time-based cross-validation and how you would prevent leakage when fitting reducers.
- Explain how you’d validate that reduction doesn’t remove critical fraud signals (including slice analysis).
- Provide a production plan: fitting cadence, monitoring, and rollback strategy.
