This document discusses approaches for fraud detection using labeled and unlabeled datasets. For labeled data, a classification model can be trained to detect fraud. For unlabeled data, an autoencoder or isolation forest can be used. An autoencoder learns a compressed representation of normal transactions and flags outliers based on reconstruction error. Isolation forest detects outliers by measuring the number of splits needed to isolate observations. The document provides examples of implementing these techniques in KNIME workflows and highlights their use in fraud detection applications.