fraud-synth-bench / docs /DATA_CARD.md
thanatchon36's picture
Add the 45 synthetic datasets, run logs and dataset card (part 2)
d632223 verified
|
Raw History Blame Contribute Delete
9.46 kB

Data Card: fraudlab.synth synthetic transaction data

Datasheet-style documentation for the generator in src/fraudlab/synth/ (version 0.1.0). Descriptive statistics for the small preset are in DATA_CARD_STATS.md, regenerated by experiments/data_card_stats.py.

This is synthetic data. No real transactions, customers or institutions were used to build, fit or calibrate the generator. Its distributions are stylised from qualitative descriptions in the published literature (see ../literature/REVIEW.md). Findings on this data are evidence about mechanisms, not about any real payment system.

Motivation

  • Purpose: controlled experiments on fraud-detection models where the fraud mechanism, its timing (drift), label latency and relational structure are known and adjustable. Real data can't offer that (REVIEW §9, gap G9).
  • Created by: the fraud-synth-bench project, Apache-2.0.
  • Not intended for: estimating real-world fraud prevalence or losses, benchmarking vendor products, or training production models.

Composition

File Grain Key columns
transactions.parquet one row per transaction tx_id, timestamp, account_id, merchant_id (null for transfers), counterparty_id (null for card payments), device_id (null for card-present), amount, mcc, channel (pos / online / transfer), country
labels.parquet one row per transaction is_fraud (ground truth), observed_label (0 if never detected), label_available_at (null unless observed), scenario, fraud_event_id
accounts.parquet one row per account opened_at, home_country, segment, account_role
merchants.parquet one row per merchant mcc, country, opened_at
edges.parquet one row per (src, dst, type) pays (account→merchant), transfers (account→account), uses_device (account→device), with first_seen, last_seen, n_tx
manifest.json per dataset generator version, preset, seed, all parameters, content hash, summary stats, library versions

Ground-truth-only columns. labels.is_fraud, labels.scenario, labels.fraud_event_id and accounts.account_role reveal the answer. They exist for evaluation and analysis, and must never be used as model features. Models may use observed_label only where label_available_at ≤ the training cutoff. The aggregated edges table spans the whole timeline, so time-respecting graph features must be built from transactions with a cutoff (Phase 4 does this).

Presets.

Preset Accounts Days ≈ Transactions Used for
tiny 2,000 30 20k tests, make quick
small 20,000 30 200k RQ1, RQ3, RQ5
small_long 8,000 90 200k RQ2, RQ4
full 200,000 30 2M documented and tested, not in the paper

On an M4 MacBook Air, small generates in about 0.2 s with a peak RSS of about 0.24 GB. full takes about 4 s with a peak RSS of about 1.2 GB.

Collection process (simulation)

Legitimate behaviour (world.py):

  • Account activity is heterogeneous: rate ~ Gamma(1.5) with mean tx_per_account_day. 3% of accounts open during the window.
  • Amounts are log-normal: an account-level location from the segment (standard / premium / student) plus noise, a category offset, and σ = 0.8.
  • Time of day follows a von Mises distribution (κ = 2.5) around each account's preferred hour (about N(14h, 3.5h)), with 10% uniform activity.
  • 65% of card payments go to one of five habitual merchants. The rest go to merchants drawn by Zipf-distributed popularity.
  • 5% of legitimate transactions are P2P transfers to one of three fixed contacts.
  • Card-present payments have no customer device. Online payments and transfers carry one.

Legitimate look-alikes, scaled by noise (default 0.4, set by the D014 difficulty check). They exist so that no single raw field separates the classes:

  • trips abroad (10% × noise of accounts, 2–7 days);
  • devices shared within a household (15% × noise);
  • switching to a new phone mid-window (10% × noise);
  • one-off logins from a fresh device (2% × noise of online transactions);
  • unusually large purchases (1% × noise, 4–15× the usual amount).

Fraud comes from eight injectors (scenarios/). Each has a docstring citing the pattern it imitates:

Scenario Mechanism Main signal Structure
card_testing 5–15 micro-charges at uniformly chosen online merchants in about 10 min; sometimes a large purchase follows tiny amounts, sub-minute gaps attacker device pool
account_takeover 2–6 high-value online purchases within 2 days from an attacker device, often abroad and at night new device, amount ≫ norm attacker device pool
velocity_abuse 8–25 purchases of $20–150 in 1–2 h at 1–3 merchants of one category count per hour individual
friendly_fraud the customer's own normal-looking purchase, later disputed almost none (by design) individual
bust_out an established account spends 3–10× its norm for 1–3 days, then goes dark change against own history individual (sequence)
mule_network victims → 5–12 young mule accounts → one collector, within hours transfers, shared devices, fan-in ring
merchant_collusion a new merchant charges 5–20 stolen cards round amounts new merchant with no legitimate customers ring
synthetic_identity 3–8 accounts opened in the window, sharing 1–2 devices, buying resellable goods young accounts, shared devices ring

Fraud budget. Within each drift segment, fraud transactions = fraud_rate × legitimate volume / (1 − fraud_rate). This is split across the active scenarios in proportion to their weights. Each injector generates whole events until its budget is met, then trims the time-ordered tail of the last event. The realised rate matches the target to within rounding (bust-out accounts going dark removes a few legitimate transactions afterwards).

Drift. A drift schedule is a list of segments over [0, 1] of the timeline, each with its own scenario weights. DRIFT_RQ2 switches from {card testing, ATO, velocity} to {bust-out, synthetic identity, mule network} at 60%. The legitimate behaviour doesn't drift.

Labels. Each fraud event is detected with probability detection_prob (0.95). The delay is drawn once per event from Uniform(7, 30) days, plus up to 2 days of per-transaction jitter. For friendly fraud it's multiplied by 1.5, since disputes are filed late. Undetected fraud keeps observed_label = 0 forever.

Determinism. Every component uses its own named RNG stream (numpy.random.default_rng([seed, crc32(name)])). The output is a pure function of the config, and adding a scenario doesn't change legitimate traffic. content_hash is a SHA-256 over column names, dtypes and values of all tables. It doesn't depend on parquet writer metadata. Same seed, same hash: this is tested, and also checked across write and load.

Validation (tests/test_synth.py)

  • Determinism: same seed gives the same hash, a different seed a different hash, and write → load preserves the hash.
  • Fraud rate within 5% (relative) of the target, at 0.1%, 0.5% and 1%.
  • Schema, temporal ordering, and the null structure of devices, merchants and counterparties.
  • Label-delay bounds, friendly-fraud delay > other fraud, detection rate.
  • The drift schedule switches the scenario set at the right time. Ring share 0 and 1 give disjoint scenario sets.
  • The noise knob adds look-alikes: card-present payments abroad, shared devices, devices per account.
  • One signature test per injector. Each passes on seeds 0–4.

Known limitations and differences from real data

  1. Not calibrated to real data. Distribution parameters are stylised. Real data has heavier tails, many more merchant categories, and seasonality (weekly or monthly cycles, holidays) that the generator doesn't model.
  2. Fraud is more mechanistic than in reality. Each scenario has a clean signature, and in practice fraudsters mix tactics. Most scenarios are easier to detect than real fraud, except friendly_fraud, which is close to undetectable by design. Absolute AUPRC values shouldn't be compared with published numbers on real data. Only relative comparisons under controlled conditions are meaningful.
  3. Few events for ring scenarios. At small with a 0.5% fraud rate, ring scenarios contribute only about 4–8 events each (about 120 transactions). Per-scenario results therefore have high variance, and results are reported over 5 seeds.
  4. No adaptive adversary. Drift is scheduled, not a reaction to the deployed model, and there's no feedback loop (FUTURE_WORK, G10).
  5. Legitimate behaviour is stationary. There are no customer life events or holiday spikes, and only fraud drifts.
  6. Single currency, abstract countries (C00–C11). There are no real geographies, merchants or identifiers, and therefore no PII.
  7. The edge table is aggregated over the whole timeline. Using it directly as a feature source leaks future information.

Maintenance

Versioned with the package (generator_version in each manifest). Any change that alters the output for a given config must bump the version and be noted in docs/DECISIONS.md.