FraudLens — Interpretable Fraud Detection for Payment Transactions

Track: Cybersecurity & Fraud Detection · DATANOVA 2026 GLOBAL DATA-THON Deliverables: Slide deck (FraudLens-DATANOVA.pptx — attached to the Devpost submission; slide content in SLIDES.md) · Working demo: https://fraudlens-eight.vercel.app · Source code: https://github.com/CloudCorpRecords/fraudlens · Optional video (walkthrough script in DEMO.md)

Overview

FraudLens is a fraud-detection pipeline and analyst dashboard that scores every payment transaction for risk and explains every score. An interpretable logistic-regression model — implemented from scratch in the Python standard library — is validated honestly against the rules-only baseline it is designed to replace, on a held-out test set, with every reported figure reproducible from artifacts included in the repository.

Problem

Payment fraud is rare (~2% of transactions), fast-moving, and expensive. Small and mid-size processors still depend on hand-written rules that miss evolving patterns (in our test set, a typical rules baseline misses 20 of 52 frauds), while black-box ML scores cannot answer the question customers, auditors, and regulators ask first: why was this transaction flagged? Fraud teams need to catch more fraud, keep false alerts low enough for their review capacity, and explain every single decision.

Solution / Methodology

1. Reproducible synthetic data. A seeded generator (seed 42) creates 8,000 transactions with a 2.14% fraud rate. Fraud is drawn from a separate behavioural process modelled on stolen-card usage: foreign country, unrecognised device, burst velocity, night hours, high-risk merchant categories, and bimodal probe/cash-out amounts. No real customer data is used.
2. Interpretable model. Logistic regression (gradient descent, class weighting, L2 regularisation) over 10 human-readable features. The coefficients are the explanation layer — the dashboard renders them as feature drivers and attaches reason data to every alert.
3. Honest validation. Stratified 70/30 split; the decision threshold is selected on the training split only and frozen before testing. A transparent points-based rules baseline is evaluated on the identical test rows. Per-row predictions are published (test_predictions.csv) so every metric can be recomputed independently.

Tech Stack

Python 3 standard library only — data generation, model training, evaluation (precision/recall/F1, ROC-AUC via Mann-Whitney). Static HTML/CSS/vanilla JavaScript dashboard — hand-rolled charts, no build step, no dependencies. CSV/JSON audit artifacts: transactions.csv, test_predictions.csv, results.json.

Implementation / Demo

One command runs the whole pipeline: python3 run.py, then python3 -m http.server 8000 --directory dashboard. The dashboard shows headline metric cards, a model-vs-rules comparison table, an interactive threshold slider that recomputes precision/recall/confusion live from the real embedded test scores, the feature-driver chart, and the flagged-transaction queue (47 alerts at the operating threshold) with the data behind each flag.

Results / Validation

Held-out test set (n = 2,401; 52 fraud / 2,349 legitimate): FraudLens model — Precision 93.6%, Recall 84.6% (44/52), F1 0.889, ROC-AUC 0.997, Confusion 44 / 3 / 2,346 / 8, Fraud value caught $14,662.78 of $16,523.91 (88.7%). Rules baseline — Precision 91.4%, Recall 61.5% (32/52), F1 0.736, ROC-AUC 0.981, Confusion 32 / 3 / 2,346 / 20, Fraud value caught $13,282.29 (80.4%). The model catches 12 more frauds than rules at the identical false-alert count (3); missed fraud falls 60% (20 to 8). Strongest drivers: burst behaviour, unrecognised device (1.91x), night activity (1.87x); account age is protective (0.60x).

Stated limitation: the dataset is synthetic with learnable patterns, so these metrics validate the method and the interpretability layer rather than predict production performance; deployment would retrain and recalibrate on real labelled transactions. Full detail, including all limitations, is in RESULTS.md in the repository.

Impact

More fraud stopped, same team: +23.1 points of recall at zero extra false alerts. Explainable decisions: every alert carries human-readable drivers — usable in customer disputes, audits, and regulatory reviews. Accessible to small teams: zero-infrastructure (stdlib Python + a static page). Responsible by construction: synthetic data only, no demographic features, and documented bias-review flags for country/distance features.

Team Contributions

Rene Turcios (solo participant) — track and problem selection, direction, and acceptance review. AI agent fleet (Tree / Muse) — dataset design and generation, model implementation, evaluation, dashboard build, and documentation, working under Rene's direction.

Build Disclosure

Build Disclosure: AI-built. Built by AI agents (Tree / Muse fleet) directed by Rene Turcios during the hackathon period; Rene set the direction and reviewed; agents designed, coded, ran, and documented the analysis.

All results in this submission are genuine outputs of the included pipeline (python3 run.py to results.json); no results are plagiarised or fabricated. Code is MIT-licensed.

Built With

• css
• html
• javascript
• python