transaction anomaly detection

problem

detect unusual card transactions that may be fraud or money laundering.
fraud is very rare (appr 0.17% of transactions).
goal was to test anomaly detection methods and compare them.

data

dataset: credit card fraud detection (europe, 2013).
284,807 transactions, 492 frauds.
features: pca components v1–v28, plus time, amount.
label: class (0 normal, 1 fraud).

methods

• z-score baseline
• isolation forest
• dbscan

all methods trained unsupervised, labels used only for evaluation.

results

summary table:

method	precision	recall	f1
zscore	0.01	0.90	0.02
isolation forest	0.15	0.44	0.23
dbscan	0.00	0.24	0.01

chart of precision vs recall is in /docs/precision_recall_bar.png.

conclusion:

• zscore finds almost all fraud but too many false alerts.
• isolation forest is more balanced.
• dbscan performs poorly here.

what to discuss in interview

• fraud is highly imbalanced, so accuracy is not useful.
• precision vs recall trade-off → false positives waste time, false negatives miss fraud.
• why anomaly detection? good when labels are scarce.
• why isolation forest worked best here.
• real world: need feature engineering (velocity, geography, peer comparison).
• real world: alerts must be manageable volume for investigators.

how to run

1. clone repo
2. install requirements

   pip install -r requirements.txt