Topological Anomaly Detection

This proposal describes a system for detecting anomalies in financial data using topological methods. It enables the efficient triage of such data, alerting a human intelligence analyst when an anomalous Persona is detected so they can be investigated further. The system models this problem as a directed multi/graph, where each node is a Persona and each edge is a Transaction directed from sender to receiver. The probability that any Persona is a bad/actor is determined by computing the persistence homology of the subgraph for which the Persona is a sink, then passing the resulting barcode to an ensemble of binary classifiers. When this probability is greater than a configured threshold, an alert is dispatched indicating that the Persona has an anomalous transaction history. This approach allows for the efficient classification of Personas using information in the transaction graph at any degree of separation.