FraudOxidizer is a Rust workspace for the Valsoft Fraud Hunter challenge. It combines:

• A model crate that parses transactions.csv, computes fraud factors, and assigns an explainable fraud score.
• An interactive egui reviewer app (desktop + web via trunk) for triage and CSV export.
• A notebook (fraud_patterns.ipynb) used to explore and validate the core fraud hypotheses.

This repository addresses the challenge requirements:

1. Ingest transactions.csv (1000 rows expected)
2. Flag suspicious transactions with a score (fraud_score)
3. Explain each flag with human-readable reasons (FraudFactor::reason())
4. Provide a reviewer workflow (approve / mark fraud, undo / redo, threshold tuning)
5. Include README and hypothesis log
6. Export an updated CSV with a likely_fraud column that respects human review decisions

The detector uses per-transaction fraud factors, then combines them with noisy-OR:

score = 1 - product(1 - weight_i)

Default fraud threshold: 0.50.

• Card Testing Burst
  • Rapid, small online transactions on the same card in a short window.
• Cashout-Like Category Risk
  • Gift card / electronics categories are treated as higher-risk contexts.
• Category Price Deviation
  • Amount is measured against category baseline using standard deviation bands (1z/2z/3z visualized in review charts).
• Card Amount Deviation
  • Amount deviates strongly from the card's own spending profile.
• Merchant Ring (cross-card)
  • Merchant with repeated high outliers across multiple distinct cards.
• Foreign Country Trip
  • Physical transactions in non-home country, with travel-aware handling to reduce false positives.
• Fraudulent Identity Link (cross-card)
  • Device/IP re-use from known suspicious transactions.
  • Confidence is stronger when linked transactions are human-confirmed fraud, weaker when only threshold-likely.

The UI is a 3-step flow:

1. Import CSV
2. Review flagged transactions
3. Overview + export reviewed CSV

Reviewer features:

• Threshold slider for human review queue
• One-by-one review carousel with contextual plots
• Keyboard shortcuts for navigation and decisions
• Bulk actions for card-testing burst series
• Undo/redo with activity log
• Combined-signal indicators in the UI
• Export that uses human decision when present (TruePositive / FalsePositive) and otherwise falls back to score threshold

Prerequisites:

• Rust toolchain (workspace uses edition 2024)
• For web UI: trunk

From repo root:

cargo run -p fraud_oxidizer

From crates/ui:

trunk serve

Then open the local URL printed by trunk.

From repo root:

cargo run -p model

This prints likely fraud transactions and their reason weights.

From repo root:

cargo fmt --all
cargo build --workspace
cargo clippy --workspace --all-targets

CI-like script:

./crates/ui/check.sh

fraud_patterns.ipynb documents and visualizes the four core challenge patterns:

1. Card Testing
2. Cashout
3. Merchant Ring
4. Impossible Travel

It also demonstrates noisy-OR combined scoring aligned with the Rust model.

Exported CSV appends a likely_fraud column.

• TruePositive review forces likely_fraud=true
• FalsePositive review forces likely_fraud=false
• Unreviewed rows use model threshold decision

• Add quantitative evaluation against labeled ground truth (precision / recall / F1 dashboard)
• Add confidence calibration and threshold tuning by explicit false-positive cost
• Improve factor-level suppression learning after reviewer feedback (beyond identity-link updates)
• Add richer audit trail export (who changed what, when, and why)
• Expand automated tests for every fraud pass and reviewer side effects