Business Context

NovaPay processes millions of card transactions per day and needs a lightweight fraud model for near-real-time scoring. The ML team wants to evaluate how activation function choice affects training stability, nonlinearity, and fraud detection performance in a feed-forward neural network.

Dataset

You are given a transaction-level binary classification dataset derived from 6 months of payment activity.

• Rows: 420K transactions, 30 input features
• Target: is_fraud (1 if confirmed fraud within 14 days, else 0)
• Class balance: 1.9% fraud, 98.1% non-fraud
• Missing data: ~6% missing in device and risk fields; categorical cardinality is moderate

Success Criteria

A strong solution should demonstrate why activation functions are necessary, compare at least three options (for example ReLU, tanh, and sigmoid in hidden layers), and achieve PR-AUC > 0.45 with recall >= 0.75 at precision >= 0.20 on the test set.

Constraints

• Inference latency must remain under 20 ms per transaction in batch-tested production conditions
• The solution should be explainable to non-research stakeholders at a practical level
• Training must fit on a single CPU or modest GPU

Deliverables

1. Explain what activation functions are and why neural networks need them.
2. Build a baseline model and compare at least three hidden-layer activation functions.
3. Describe preprocessing for numerical, categorical, and missing features.
4. Evaluate models using fraud-appropriate metrics and choose a deployment candidate.
5. Discuss training stability, threshold selection, and production tradeoffs.