Skip to main content

Fraud detection

A fraud model is only as good as its features. With Feldera, you write the features once in SQL. The same queries compute training data from history and compute features for each live transaction, so training and scoring always agree.

Transactionshistory + live streamFeature SQLjoins, rolling windowsModeltraining and scoringevery changefeature vectors
One set of queries for training and inference. Perfect parity, no matter how complex.

Why Feldera​

  • One set of queries. Train on Delta Lake history and score live transactions from Kafka with the same SQL. You change only the connectors. The features are identical in training and inference, no matter how complex the queries.
  • Exact features for each transaction. Joins and rolling aggregates are updated incrementally, so each feature vector is correct when the transaction arrives.
  • Models in the loop. A model can read its inputs from a view and write its scores to a table, so the pipeline measures its accuracy while it runs.

Guides and demos​

GuideReal-time fraud featuresWrite fraud features in SQL, train a model on historical data, then score a live transaction stream.GuideKeeping an ML model in the loopFeed a model from a view, read its predictions from a table, and score its accuracy as data arrives.DemoFraud detection on Delta LakeCompute features from public Delta tables on S3, train an XGBoost model, then score in real time.DemoHopsworks feature storeFeldera computes features from Hopsworks Kafka. Hopsworks stores them and serves the model.DemoAgentic fraud detectionAn agent reads a report about a new fraud pattern and writes SQL views to detect it.